Skip to content

2
2026_Article_RESILINK_ML
Commit 1a8000f5 by Joaquín Irazábal González
Options

Refactor code structure for improved readability and maintainability

parent d3050c70
Showing with 47 additions and 25 deletions
Code/reports/width_optimization/3W/ml_optimization/it3/optimization_results_3W_B29_H45_TFD_W100.csv
......@@ -7,6 +7,15 @@ Iteration,3.0
tw1_optimal,6.736822597759505
tw2_optimal,8.8630649168025
tw3_optimal,10.176584490906917
candidate_tw1_raw,6.736822597759505
candidate_tw1,6.74
rounding_delta_tw1,0.0031774022404951197
candidate_tw2_raw,8.8630649168025
candidate_tw2,8.86
rounding_delta_tw2,-0.003064916802500761
candidate_tw3_raw,10.176584490906917
candidate_tw3,10.18
rounding_delta_tw3,0.003415509093082747
Objective_score,-1.0833005004210855e-06
Exy_tw1,0.041973753876904246
Exy_tw2,0.045355043140069135
......@@ -15,3 +24,10 @@ TFM_tw1,73.67731201163258
TFM_tw2,73.67731208919551
TFM_tw3,73.6821569199441
TFM_frame,60.074381369440665
manuf_Exy_tw1,0.041946632475445694
manuf_Exy_tw2,0.045385886970777306
manuf_Exy_tw3,0.04771138851832256
manuf_TFM_tw1,73.60490265248133
manuf_TFM_tw2,73.74469924416302
manuf_TFM_tw3,73.64764678728832
manuf_TFM_frame,60.0713624885057
Code/reports/width_optimization/3W/ml_optimization/ml_3W_B29_H45_TFD_W100_it3_lexicographic_optimization_summary.csv
w_val,b_val,h_val,tfd_w_val,it,safety_margin_mode,safety_margin_factor,local_margin_support_pass,local_margin_k_requested,local_support_percentile,local_margin_fallback,TFD_tw1_margin_source,TFD_tw1_margin_in_support,TFD_tw1_margin_k,TFD_tw1_margin_nearest_distance,TFD_tw1_margin_support_radius,TFD_tw1_margin_neighbor_samples,TFD_tw1_margin_neighbor_distances,TFD_tw1_margin_neighbor_underprediction,TFD_tw2_margin_source,TFD_tw2_margin_in_support,TFD_tw2_margin_k,TFD_tw2_margin_nearest_distance,TFD_tw2_margin_support_radius,TFD_tw2_margin_neighbor_samples,TFD_tw2_margin_neighbor_distances,TFD_tw2_margin_neighbor_underprediction,TFD_tw3_margin_source,TFD_tw3_margin_in_support,TFD_tw3_margin_k,TFD_tw3_margin_nearest_distance,TFD_tw3_margin_support_radius,TFD_tw3_margin_neighbor_samples,TFD_tw3_margin_neighbor_distances,TFD_tw3_margin_neighbor_underprediction,TFD_frame_margin_source,TFD_frame_margin_in_support,TFD_frame_margin_k,TFD_frame_margin_nearest_distance,TFD_frame_margin_support_radius,TFD_frame_margin_neighbor_samples,TFD_frame_margin_neighbor_distances,TFD_frame_margin_neighbor_underprediction,min_de_reproducibility_required,stage1_performance_reproducibility_rate,stage2_performance_reproducibility_rate,stage1_performance_reproducibility_pass,stage2_performance_reproducibility_pass,de_performance_reproducibility_pass,stage1_geometry_reproducibility_rate,stage2_geometry_reproducibility_rate,stage1_geometry_n_basins,stage2_geometry_n_basins,geometry_unique_or_reproducible,stage1_performance_spread_min_window_tfd,stage2_best_distortion_measure,stage2_performance_spread_distortion,feasible_domain_found,minimum_domain_violation,feasible,tw1,tw2,tw3,TFD_tw1,TFD_tw2,TFD_tw3,TFD_tw1_margin,TFD_tw2_margin,TFD_tw3_margin,TFD_tw1_conservative,TFD_tw2_conservative,TFD_tw3_conservative,TFD_frame,TFD_frame_margin,TFD_frame_conservative,min_window_tfd,max_window_tfd,mean_window_tfd,window_tfd_range_diagnostic,distortion_measure,stage1_Tmin_star,stage1_preservation_threshold,max_constraint_violation,stage1_preservation_tolerance,stage1_preservation_tolerance_source,stage1_performance_group_count,stage1_rep_run,stage1_rep_seed,stage2_rep_run,stage2_rep_seed,stage2_tw1_dominant_basin_iqr,stage2_tw2_dominant_basin_iqr,stage2_tw3_dominant_basin_iqr,stage2_tw1_dominant_basin_range,stage2_tw2_dominant_basin_range,stage2_tw3_dominant_basin_range
3,29,45,100,3,local_oof,1.0,True,-1,95.0,global,local_knn_max,True,5,0.07775922294967777,0.5773528928252006,16;18;17;2;0,0.077759223;0.12843261;0.14166216;0.2737436;0.37163663,5.6201407;0;3.8305683;0;7.6020582,local_knn_max,True,5,0.07775922294967777,0.5773528928252006,16;18;17;2;0,0.077759223;0.12843261;0.14166216;0.2737436;0.37163663,0;1.9587951;0;4.7321588;8.5168892,local_knn_max,True,5,0.07775922294967777,0.5773528928252006,16;18;17;2;0,0.077759223;0.12843261;0.14166216;0.2737436;0.37163663,0;3.4272786;0;9.5428124;0,local_knn_max,True,5,0.07775922294967777,0.5773528928252006,16;18;17;2;0,0.077759223;0.12843261;0.14166216;0.2737436;0.37163663,0.048146252;0.21810102;0;1.7317462;0,0.8,1.0,1.0,True,True,True,1.0,1.0,1,1,True,3.7058609606788195e-05,1.0833005004210855e-06,4.0500712956902346e-13,True,0.0,True,6.736822597759505,8.8630649168025,10.176584490906917,73.67731201163258,73.67731208919551,73.6821569199441,7.602058219372225,8.516889231763038,9.542812368175888,81.2793702310048,82.19420132095854,83.22496928812,60.074381369440665,1.7317461805155503,61.806127549956216,73.67731201163258,73.6821569199441,73.67892700692407,0.004844908311525842,1.0833005004210855e-06,73.67831201092002,73.67731201092002,0.0,0.001,performance_group_spread,30,12,54,25,67,1.0487073955900428e-06,6.647890362643238e-07,5.906639270136793e-06,3.740404444840806e-06,2.677528186723066e-06,2.0328844867023577e-05
w_val,b_val,h_val,tfd_w_val,it,safety_margin_mode,safety_margin_factor,local_margin_support_pass,local_margin_k_requested,local_support_percentile,local_margin_fallback,TFD_tw1_margin_source,TFD_tw1_margin_in_support,TFD_tw1_margin_k,TFD_tw1_margin_nearest_distance,TFD_tw1_margin_support_radius,TFD_tw1_margin_neighbor_samples,TFD_tw1_margin_neighbor_distances,TFD_tw1_margin_neighbor_underprediction,TFD_tw2_margin_source,TFD_tw2_margin_in_support,TFD_tw2_margin_k,TFD_tw2_margin_nearest_distance,TFD_tw2_margin_support_radius,TFD_tw2_margin_neighbor_samples,TFD_tw2_margin_neighbor_distances,TFD_tw2_margin_neighbor_underprediction,TFD_tw3_margin_source,TFD_tw3_margin_in_support,TFD_tw3_margin_k,TFD_tw3_margin_nearest_distance,TFD_tw3_margin_support_radius,TFD_tw3_margin_neighbor_samples,TFD_tw3_margin_neighbor_distances,TFD_tw3_margin_neighbor_underprediction,TFD_frame_margin_source,TFD_frame_margin_in_support,TFD_frame_margin_k,TFD_frame_margin_nearest_distance,TFD_frame_margin_support_radius,TFD_frame_margin_neighbor_samples,TFD_frame_margin_neighbor_distances,TFD_frame_margin_neighbor_underprediction,min_de_reproducibility_required,stage1_performance_reproducibility_rate,stage2_performance_reproducibility_rate,stage1_performance_reproducibility_pass,stage2_performance_reproducibility_pass,de_performance_reproducibility_pass,stage1_geometry_reproducibility_rate,stage2_geometry_reproducibility_rate,stage1_geometry_n_basins,stage2_geometry_n_basins,geometry_unique_or_reproducible,stage1_performance_spread_min_window_tfd,stage2_best_distortion_measure,stage2_performance_spread_distortion,feasible_domain_found,minimum_domain_violation,feasible,tw1,tw2,tw3,geometry_projection_applied,manufacturing_thickness_resolution_mm,candidate_tw1_raw,candidate_tw2_raw,candidate_tw3_raw,candidate_tw1,candidate_tw2,candidate_tw3,rounding_delta_tw1,rounding_delta_tw2,rounding_delta_tw3,manuf_min_window_tfd,manuf_max_window_tfd,manuf_mean_window_tfd,manuf_distortion_measure,manuf_frame_tfd,manuf_frame_tfd_conservative,manuf_max_constraint_violation,manuf_feasible,manuf_local_margin_support_pass,manuf_TFD_tw1,manuf_TFD_tw2,manuf_TFD_tw3,manuf_TFD_tw1_conservative,manuf_TFD_tw2_conservative,manuf_TFD_tw3_conservative,TFD_tw1,TFD_tw2,TFD_tw3,TFD_tw1_margin,TFD_tw2_margin,TFD_tw3_margin,TFD_tw1_conservative,TFD_tw2_conservative,TFD_tw3_conservative,TFD_frame,TFD_frame_margin,TFD_frame_conservative,min_window_tfd,max_window_tfd,mean_window_tfd,window_tfd_range_diagnostic,distortion_measure,stage1_Tmin_star,stage1_preservation_threshold,max_constraint_violation,stage1_preservation_tolerance,stage1_preservation_tolerance_source,stage1_performance_group_count,stage1_rep_run,stage1_rep_seed,stage2_rep_run,stage2_rep_seed,stage2_tw1_dominant_basin_iqr,stage2_tw2_dominant_basin_iqr,stage2_tw3_dominant_basin_iqr,stage2_tw1_dominant_basin_range,stage2_tw2_dominant_basin_range,stage2_tw3_dominant_basin_range
3,29,45,100,3,local_oof,1.0,True,-1,95.0,global,local_knn_max,True,5,0.07775922294967777,0.5773528928252006,16;18;17;2;0,0.077759223;0.12843261;0.14166216;0.2737436;0.37163663,5.6201407;0;3.8305683;0;7.6020582,local_knn_max,True,5,0.07775922294967777,0.5773528928252006,16;18;17;2;0,0.077759223;0.12843261;0.14166216;0.2737436;0.37163663,0;1.9587951;0;4.7321588;8.5168892,local_knn_max,True,5,0.07775922294967777,0.5773528928252006,16;18;17;2;0,0.077759223;0.12843261;0.14166216;0.2737436;0.37163663,0;3.4272786;0;9.5428124;0,local_knn_max,True,5,0.07775922294967777,0.5773528928252006,16;18;17;2;0,0.077759223;0.12843261;0.14166216;0.2737436;0.37163663,0.048146252;0.21810102;0;1.7317462;0,0.8,1.0,1.0,True,True,True,1.0,1.0,1,1,True,3.7058609606788195e-05,1.0833005004210855e-06,4.0500712956902346e-13,True,0.0,True,6.736822597759505,8.8630649168025,10.176584490906917,True,0.01,6.736822597759505,8.8630649168025,10.176584490906917,6.74,8.86,10.18,0.0031774022404951197,-0.003064916802500761,0.003415509093082747,73.60490265248133,73.74469924416302,73.66574956131089,1.0835459118768507e-06,60.0713624885057,61.80310866902125,0.0,True,True,73.60490265248133,73.74469924416302,73.64764678728832,81.20696087185355,82.26158847592606,83.19045915546421,73.67731201163258,73.67731208919551,73.6821569199441,7.602058219372225,8.516889231763038,9.542812368175888,81.2793702310048,82.19420132095854,83.22496928812,60.074381369440665,1.7317461805155503,61.806127549956216,73.67731201163258,73.6821569199441,73.67892700692407,0.004844908311525842,1.0833005004210855e-06,73.67831201092002,73.67731201092002,0.0,0.001,performance_group_spread,30,12,54,25,67,1.0487073955900428e-06,6.647890362643238e-07,5.906639270136793e-06,3.740404444840806e-06,2.677528186723066e-06,2.0328844867023577e-05
Code/reports/width_optimization/5W/rbf_optimization/rbf_5W_B34_H60_TFD_W100_it2_adaptive_acceptance_summary.csv
surrogate,W,B,H,TFD_W_limit,TFD_frame_limit,iteration,candidate_row_zero_based,geometry_match,geometry_match_tolerance_mm,geometry_projection_applied,manufacturing_thickness_resolution_mm,surrogate_validation_geometry_matches_fem_candidate,tfd_physical_acceptance_resolution,tfd_physical_acceptance_decimals,tfd_physical_acceptance_rounding,fem_tfmmax_tw1_raw,fem_tfmmax_tw1_acceptance,fem_tfmmax_tw1_limit,fem_tfmmax_tw1_physical_pass,fem_tfmmax_tw2_raw,fem_tfmmax_tw2_acceptance,fem_tfmmax_tw2_limit,fem_tfmmax_tw2_physical_pass,fem_tfmmax_tw3_raw,fem_tfmmax_tw3_acceptance,fem_tfmmax_tw3_limit,fem_tfmmax_tw3_physical_pass,fem_tfmmax_tw4_raw,fem_tfmmax_tw4_acceptance,fem_tfmmax_tw4_limit,fem_tfmmax_tw4_physical_pass,fem_tfmmax_tw5_raw,fem_tfmmax_tw5_acceptance,fem_tfmmax_tw5_limit,fem_tfmmax_tw5_physical_pass,fem_tfmmax_frame_raw,fem_tfmmax_frame_acceptance,fem_tfmmax_frame_limit,fem_tfmmax_frame_physical_pass,fem_feasible,damage_local_accuracy_pass,min_window_tfd_local_accuracy_pass,distortion_measure_local_accuracy_pass,surrogate_local_accuracy_pass,local_acceptance_support_pass,local_margin_support_pass,C_FEM_fem_feasibility,C_FEM_surrogate_local_accuracy,C_FEM_local_support,C_FEM,C_FEM_definition,local_oof_support_pass,surrogate_local_validity_pass,safety_margin_mode,de_reproducibility_warning_threshold,de_final_reproducibility_requirement,de_reproducibility_source,stage1_performance_reproducibility_rate,stage2_performance_reproducibility_rate,stage1_R_perf,stage2_R_perf,stage1_n_runs,stage2_n_runs,stage1_n_equivalent,stage2_n_equivalent,stage1_reproducibility_warning_pass,stage2_reproducibility_warning_pass,stage1_final_reproducibility_pass,stage2_final_reproducibility_pass,de_performance_reproducibility_pass,stage1_geometry_reproducibility_rate,stage2_geometry_reproducibility_rate,geometry_reproducibility_diagnostic,accepted,final_converged,recommended_action,diagnostic_recommendation,continuation_reason,failed_criteria,failed_subcriteria,diagnostic_warnings,C_DE,C_STABILITY,current_stage1_Tmin_star,previous_stage1_Tmin_star,delta_stage1_optimum,delta_stage1_optimum_percent,stage1_optimum_stable,stage1_optimum_stability_status,current_stage2_distortion,previous_stage2_distortion,delta_stage2_optimum,delta_stage2_optimum_percent,stage2_optimum_stable,stage2_optimum_stability_status,adaptive_stability_limit,adaptive_stability_limit_percent,report_schema_version,convergence_logic_version,fem_min_window_tfd,pred_min_window_tfd,fem_distortion_measure,pred_distortion_measure,surrogate_manuf_window_tfd_pass,surrogate_manuf_frame_tfd_pass,surrogate_manuf_conservative_pass,surrogate_manuf_stage1_preservation_pass,surrogate_manuf_feasibility_pass,candidate_tw1_raw,candidate_tw2_raw,candidate_tw3_raw,candidate_tw4_raw,candidate_tw5_raw,candidate_tw1,candidate_tw2,candidate_tw3,candidate_tw4,candidate_tw5,rounding_delta_tw1,rounding_delta_tw2,rounding_delta_tw3,rounding_delta_tw4,rounding_delta_tw5,delta_tw1,de_repeatability_range_tw1,delta_within_de_repeatability_tw1,delta_tw2,de_repeatability_range_tw2,delta_within_de_repeatability_tw2,delta_tw3,de_repeatability_range_tw3,delta_within_de_repeatability_tw3,delta_tw4,de_repeatability_range_tw4,delta_within_de_repeatability_tw4,delta_tw5,de_repeatability_range_tw5,delta_within_de_repeatability_tw5
rbf,5,34,60,100.0,90.0,2,66,True,0.01,True,0.01,True,0.1,1,ROUND_HALF_UP,79.6936,79.7,100.0,True,80.6339,80.6,100.0,True,79.5546,79.6,100.0,True,79.1653,79.2,100.0,True,79.455,79.5,100.0,True,69.4714,69.5,90.0,True,True,True,True,True,True,True,True,True,True,True,True,physical_fem_feasibility_only,True,True,local_oof,0.8,1.0,optimizer_summary_performance_fields,1.0,1.0,1.0,1.0,30,30,30,30,True,True,True,True,True,1.0,1.0,PERFORMANCE_AND_GEOMETRY_REPRODUCIBLE,False,False,CONTINUE_ADAPTIVE_ENRICHMENT_OPTIMUM_NOT_STABLE,,Optimum not stabilized across iterations,C_STABILITY,stage2_optimum_stability,,True,False,80.92166891936199,79.37191804662143,0.01915124704465597,1.915124704465597,True,STABLE,1.912078346172097e-06,1.958298890582856e-06,0.024172934390105195,2.4172934390105194,False,UNSTABLE,0.02,2.0,4,three_criteria_fem_feasibility_v2,79.1653,80.81461036542964,1.8464881228943121e-06,1.9104434180197957e-06,True,True,True,False,False,6.195308486124487,7.847207271585908,8.936474407844969,7.068562987823455,5.000000048625363,6.2,7.85,8.94,7.07,5.0,0.0046915138755130315,0.0027927284140920605,0.0035255921550305658,0.0014370121765452382,-4.862536329142131e-08,0.003096562483373866,0.0002108767428454,False,0.13027312022175774,0.0013067067754146,False,0.4623869703259036,0.001957019641491,False,0.17599625623775061,0.000691252654299,False,2.686984128530412e-07,0.0004203240008227,True
surrogate,W,B,H,TFD_W_limit,TFD_frame_limit,iteration,candidate_row_zero_based,geometry_match,geometry_match_tolerance_mm,geometry_projection_applied,manufacturing_thickness_resolution_mm,surrogate_validation_geometry_matches_fem_candidate,tfd_physical_acceptance_resolution,tfd_physical_acceptance_decimals,tfd_physical_acceptance_rounding,fem_tfmmax_tw1_raw,fem_tfmmax_tw1_acceptance,fem_tfmmax_tw1_limit,fem_tfmmax_tw1_physical_pass,fem_tfmmax_tw2_raw,fem_tfmmax_tw2_acceptance,fem_tfmmax_tw2_limit,fem_tfmmax_tw2_physical_pass,fem_tfmmax_tw3_raw,fem_tfmmax_tw3_acceptance,fem_tfmmax_tw3_limit,fem_tfmmax_tw3_physical_pass,fem_tfmmax_tw4_raw,fem_tfmmax_tw4_acceptance,fem_tfmmax_tw4_limit,fem_tfmmax_tw4_physical_pass,fem_tfmmax_tw5_raw,fem_tfmmax_tw5_acceptance,fem_tfmmax_tw5_limit,fem_tfmmax_tw5_physical_pass,fem_tfmmax_frame_raw,fem_tfmmax_frame_acceptance,fem_tfmmax_frame_limit,fem_tfmmax_frame_physical_pass,fem_feasible,damage_local_accuracy_pass,min_window_tfd_local_accuracy_pass,distortion_measure_local_accuracy_pass,surrogate_local_accuracy_pass,local_acceptance_support_pass,local_margin_support_pass,C_FEM_fem_feasibility,C_FEM_surrogate_local_accuracy,C_FEM_local_support,C_FEM,C_FEM_definition,local_oof_support_pass,surrogate_local_validity_pass,safety_margin_mode,de_reproducibility_warning_threshold,de_final_reproducibility_requirement,de_reproducibility_source,stage1_performance_reproducibility_rate,stage2_performance_reproducibility_rate,stage1_R_perf,stage2_R_perf,stage1_n_runs,stage2_n_runs,stage1_n_equivalent,stage2_n_equivalent,stage1_reproducibility_warning_pass,stage2_reproducibility_warning_pass,stage1_final_reproducibility_pass,stage2_final_reproducibility_pass,de_performance_reproducibility_pass,stage1_geometry_reproducibility_rate,stage2_geometry_reproducibility_rate,geometry_reproducibility_diagnostic,accepted,final_converged,recommended_action,diagnostic_recommendation,continuation_reason,failed_criteria,failed_subcriteria,diagnostic_warnings,C_DE,C_STABILITY,C_CONS,conservatism_gap_rtol,conservatism_gap_eps,stage1_Tmin_star_conservative,stage1_Tmin_star_shadow,stage1_conservatism_gap_absolute,stage1_conservatism_gap_relative,stage1_conservative_margin_materially_limiting,stage1_shadow_status,shadow_stage1_assessable,shadow_stage1_n_runs,shadow_stage1_seed_base,shadow_stage1_equivalent_runs,shadow_stage1_R_perf,shadow_stage1_geometry_reproducibility,shadow_stage1_max_constraint_violation,stage1_shadow_message,shadow_stage1_tw1,shadow_stage1_tw2,shadow_stage1_tw3,shadow_stage1_tw4,shadow_stage1_tw5,current_stage1_Tmin_star,previous_stage1_Tmin_star,delta_stage1_optimum,delta_stage1_optimum_percent,stage1_optimum_stable,stage1_optimum_stability_status,current_stage2_distortion,previous_stage2_distortion,delta_stage2_optimum,delta_stage2_optimum_percent,stage2_optimum_stable,stage2_optimum_stability_status,adaptive_stability_limit,adaptive_stability_limit_percent,report_schema_version,convergence_logic_version,fem_min_window_tfd,pred_min_window_tfd,fem_distortion_measure,pred_distortion_measure,surrogate_manuf_window_tfd_pass,surrogate_manuf_frame_tfd_pass,surrogate_manuf_conservative_pass,surrogate_manuf_stage1_preservation_pass,surrogate_manuf_feasibility_pass,candidate_tw1_raw,candidate_tw2_raw,candidate_tw3_raw,candidate_tw4_raw,candidate_tw5_raw,candidate_tw1,candidate_tw2,candidate_tw3,candidate_tw4,candidate_tw5,rounding_delta_tw1,rounding_delta_tw2,rounding_delta_tw3,rounding_delta_tw4,rounding_delta_tw5,delta_tw1,de_repeatability_range_tw1,delta_within_de_repeatability_tw1,delta_tw2,de_repeatability_range_tw2,delta_within_de_repeatability_tw2,delta_tw3,de_repeatability_range_tw3,delta_within_de_repeatability_tw3,delta_tw4,de_repeatability_range_tw4,delta_within_de_repeatability_tw4,delta_tw5,de_repeatability_range_tw5,delta_within_de_repeatability_tw5
rbf,5,34,60,100.0,90.0,2,66,True,0.01,True,0.01,True,0.1,1,ROUND_HALF_UP,79.6936,79.7,100.0,True,80.6339,80.6,100.0,True,79.5546,79.6,100.0,True,79.1653,79.2,100.0,True,79.455,79.5,100.0,True,69.4714,69.5,90.0,True,True,True,True,True,True,True,True,True,True,True,True,physical_fem_feasibility_only,True,True,local_oof,0.8,1.0,optimizer_summary_performance_fields,1.0,1.0,1.0,1.0,30,30,30,30,True,True,True,True,True,1.0,1.0,PERFORMANCE_AND_GEOMETRY_REPRODUCIBLE,False,False,CONTINUE_ADAPTIVE_ENRICHMENT_OPTIMUM_NOT_STABLE,,Optimum not stabilized across iterations,C_STABILITY,stage2_optimum_stability,,True,False,True,0.02,1e-12,80.92166891936199,80.92166337830713,0.0,0.0,False,OK,True,30,42,30,1.0,1.0,0.0,,6.195723992684709,7.846961953681928,8.936498665797227,7.068566983710585,5.000002080235616,80.92166891936199,79.37191804662143,0.01915124704465597,1.915124704465597,True,STABLE,1.912078346172097e-06,1.958298890582856e-06,0.024172934390105195,2.4172934390105194,False,UNSTABLE,0.02,2.0,5,four_criteria_fem_de_stability_conservatism_v3,79.1653,80.81461036542964,1.8464881228943121e-06,1.9104434180197957e-06,True,True,True,False,False,6.195308486124487,7.847207271585908,8.936474407844969,7.068562987823455,5.000000048625363,6.2,7.85,8.94,7.07,5.0,0.0046915138755130315,0.0027927284140920605,0.0035255921550305658,0.0014370121765452382,-4.862536329142131e-08,0.003096562483373866,0.0002108767428454,False,0.13027312022175774,0.0013067067754146,False,0.4623869703259036,0.001957019641491,False,0.17599625623775061,0.000691252654299,False,2.686984128530412e-07,0.0004203240008227,True
ManuscriptR1/ComparisonSurrogatesOptimizationBDSL_R1.tex
......@@ -66,12 +66,12 @@
country={Spain}}
\begin{abstract}
Buckling-delayed shear-link dampers are passive devices used in seismic-resistant structures to concentrate energy dissipation while limiting damage to the primary system. Optimizing their window thicknesses requires balancing high dissipative capacity with strict control of local damage. Nonlinear finite element models can reproduce their cyclic response and provide internal quantities for optimization, such as damage indicators and local distortion, but their computational cost prevents direct use within iterative optimization loops. This work proposes an adaptive surrogate-assisted optimization framework for damage-aware window-thickness optimization of buckling-delayed shear-link dampers within five predefined geometry families. Experimentally calibrated nonlinear finite element models are first used to generate reference datasets for different device configurations. Supervised learning models are evaluated, with support vector regression and Gaussian process regression showing high predictive accuracy, motivating the use of radial basis function surrogates as a more efficient distance-based alternative. The surrogate predictions are coupled with a differential evolution algorithm through a damage-aware objective function that controls local damage while using dissipative performance as an optimization criterion. The optimized geometries are then re-evaluated with finite element simulations. If the acceptance criteria are not met, the new simulation is added to the dataset and the surrogate models are retrained. The framework enables efficient damage-aware optimization of seismic energy dissipation devices.
Buckling-delayed shear-link dampers are passive devices used in seismic-resistant structures to concentrate energy dissipation while limiting damage to the primary system. Optimizing their window thicknesses requires balancing high dissipative capacity with strict control of local damage. Nonlinear finite element models can reproduce their cyclic response and provide internal quantities for optimization, such as damage indicators and local distortion, but their computational cost prevents direct use within iterative optimization loops. This work proposes an adaptive surrogate-assisted optimization framework for damage-aware window-thickness optimization of buckling-delayed shear-link dampers within five predefined geometry families. Experimentally calibrated nonlinear finite element models are first used to generate reference datasets for different device configurations. Supervised learning models are evaluated, with support vector regression and Gaussian process regression providing the most accurate predictions among the candidate models, motivating the use of radial basis function surrogates as a more efficient distance-based alternative. The surrogate predictions are coupled with a differential evolution algorithm through a damage-aware objective function that controls local damage while using dissipative performance as an optimization criterion. The optimized geometries are then re-evaluated with finite element simulations. If the acceptance criteria are not met, the new simulation is added to the dataset and the surrogate models are retrained. The framework enables efficient damage-aware optimization of seismic energy dissipation devices.
\end{abstract}
\begin{highlights}
\item Adaptive surrogates optimize damage-aware shear-link damper geometries
\item RBF interpolation matches supervised ML with lower retraining cost
\item RBF interpolation offers comparable accuracy to supervised ML at lower training cost
\item FEM validation filters surrogate optima before accepting final designs
\item Optimized devices balance window activation and frame damage control
\end{highlights}
......@@ -332,7 +332,7 @@ The surrogate-optimized geometry is not accepted directly. Instead, once an opti
\section{Numerical results and discussion}\label{sec:results}
The supervised-learning comparison shows a clear hierarchy among the candidate surrogate models. A total of 100 output-specific training problems were considered, corresponding to the $2N_w+1$ target variables required for each geometry family and adaptive iteration. The $F_1$ family required two optimization iterations, whereas the remaining families required three. Across all geometry families, iterations and output variables, SVR was the most frequently selected model and also the model that most often achieved the lowest cross-validated RMSE. As summarized in Figure~\ref{fig:surrogate_selection_summary_barplot}, SVR was selected in 71 cases, followed by GPR, GBR, XGBoost and MLP, while Random Forest was not selected in any case. This dominance was particularly clear for the two-window families, where only one output was assigned to a GPR model and for the damage-related outputs, for which SVR was selected in 47 out of 57 cases. For more complex devices, those with three and five windows, and for distortion-related outputs, the model selection became more heterogeneous, although SVR still provided the best overall performance. From the computational point of view, SVR also offered the lowest median training times among the supervised models, whereas MLP required substantially longer training times without providing gain in accuracy.
The supervised-learning comparison shows a predominance of kernel-based models among the candidate surrogates. A total of 100 output-specific model-selection problems were considered, corresponding to the $2N_w+1$ target variables required for each geometry family and adaptive iteration. The $F_1$ family required two optimization iterations, whereas the remaining families required three. Across all geometry families, iterations and output variables, SVR was the most frequently selected model and also the model that most often achieved the lowest cross-validated RMSE. As summarized in Figure~\ref{fig:surrogate_selection_summary_barplot}, SVR was selected in 71 cases, followed by GPR (15), GBR (10), XGBoost (2) and MLP (2), while Random Forest was not selected in any case. This predominance was particularly clear for the two-window families, where only one output was assigned to a GPR model, and for the damage-related outputs, for which SVR was selected in 47 out of 57 cases. For more complex devices, those with three and five windows, and for distortion-related outputs, the model selection became more heterogeneous, although SVR remained the most frequently selected model. The RMSE-dispersion selection criterion caused the selected model to differ from the single lowest-RMSE model only in a minority of the aggregate counts; for example, GPR attained the lowest RMSE in 17 outputs but was selected in 15, whereas GBR was selected in 10 and attained the lowest RMSE in 9. These counts are descriptive statistics over model-selection tasks that are not mutually independent, because several outputs belong to the same geometry family, adaptive iteration and correlated FEM response. They are therefore reported as a descriptive summary of the comparison rather than as independent statistical evidence of superiority, and the sensitivity of the selection to the adopted 5\% competitive RMSE threshold has not been assessed. From the computational point of view, SVR also offered the lowest median training times among the supervised models, whereas MLP required substantially longer training times without providing gain in accuracy.
\begin{figure*}[htbp]
\centering
......@@ -341,9 +341,11 @@ The supervised-learning comparison shows a clear hierarchy among the candidate s
\label{fig:surrogate_selection_summary_barplot}
\end{figure*}
These results indicate that kernel-based models are particularly well suited to the present surrogate task. SVR provides the best compromise between accuracy and computational cost, while GPR is the second most competitive supervised strategy, especially in some higher-dimensional cases. Tree-based models, although robust, are less frequently selected and MLP models are not competitive in terms of computational efficiency for the dataset sizes considered here. As mentioned in previous sections, this behaviour motivated the additional evaluation of RBF interpolation as a simpler surrogate alternative. In contrast to the supervised models, RBF models were trained in less than one second per output, making them especially attractive for repeated surrogate updates within the adaptive optimization loop.
These results indicate that kernel-based models performed well under the adopted nested validation procedure for the datasets considered here. SVR provided the best compromise between accuracy and computational cost, while GPR was the second most frequently selected supervised strategy, especially in some higher-dimensional cases. Tree-based models were less frequently selected and MLP models were not competitive in terms of computational efficiency for the dataset sizes considered. As mentioned in previous sections, this behaviour motivated the additional evaluation of RBF interpolation as a simpler surrogate alternative. Because the RBF and the supervised models are assessed on the same outer cross-validation splits and with the same error metrics, the two strategies can be compared on a like-for-like basis at the level of each output variable. Across the analysed outputs neither strategy dominates: the selected supervised model is more accurate for some targets and the RBF surrogate for others, and both remain within the accuracy range required by the adaptive validation loop. The RBF surrogate, however, is obtained with a much smaller hyperparameter search and therefore requires substantially lower training effort, which makes it attractive for repeated surrogate updates within the adaptive optimization loop.
The FEM validation of the optimized geometries is summarized in Table~\ref{tab:final_surrogate_comparison}, while the complete optimization results for all adaptive iterations are provided in Appendix~\ref{appendix:optimization_results}. For each geometry family and surrogate strategy, Table~\ref{tab:final_surrogate_comparison} reports the final accepted adaptive iteration, the optimized window thicknesses $\mathbf{t}_w^{\star}$, the surrogate-predicted objective value $J_{\mathrm{surr}}$, the corresponding FEM-recomputed objective value $J_{\mathrm{FEM}}$ and the associated validation errors, $|e_J|$ and $e_{\max}$. The maximum variable error, $e_{\max}$, is defined as the largest relative error among all quantities entering the objective function, namely ${\Exy}_i$, $\TFD_i$ and $\TFD_f$, whereas $|e_J|$ denotes the absolute objective-function error. The optimization required between two and three adaptive iterations depending on the geometry family and surrogate type, with most cases converging after three iterations. No systematic difference in the number of iterations was observed between RBF and supervised ML surrogates.
The predictive assessment is constrained by the small datasets available for the two-window families ($N=8$) and the three-window families ($N=16$). Nested cross-validation separates hyperparameter selection from generalization assessment, and the outer folds provide out-of-sample predictions together with bootstrap confidence intervals for RMSE and MAE. For datasets with $N\leq20$ the outer loop is Leave-One-Out, so the reported errors are obtained from a single deterministic partition and their dispersion is estimated by resampling the out-of-sample predictions rather than by repeating the outer split. These results should therefore be interpreted as an assessment of predictive accuracy within the sampled design domain and for the analysed geometry families, under the adopted validation procedure, rather than as a general guarantee. No independent FEM test set was retained, and the uncertainty of the accuracy estimates for the smallest datasets remains a limitation of the present comparison.
The FEM validation of the optimized geometries is summarized in Table~\ref{tab:final_surrogate_comparison}, while the complete optimization results for all adaptive iterations are provided in Appendix~\ref{appendix:optimization_results}. For each geometry family and surrogate strategy, Table~\ref{tab:final_surrogate_comparison} reports the final accepted adaptive iteration, the optimized window thicknesses $\mathbf{t}_w^{\star}$, the surrogate-predicted objective value $J_{\mathrm{surr}}$, the corresponding FEM-recomputed objective value $J_{\mathrm{FEM}}$ and the associated validation errors, $|e_J|$ and $e_{\max}$. The maximum variable error, $e_{\max}$, is defined as the largest relative error among all quantities entering the objective function, namely ${\Exy}_i$, $\TFD_i$ and $\TFD_f$, whereas $|e_J|$ denotes the absolute objective-function error. It should be emphasised that these errors measure the agreement between the surrogate and the FEM model at the optimized candidate; they characterise the local consistency of the surrogate in the vicinity of the optimum and are not, by themselves, a measure of global predictive accuracy over the whole design domain. The optimization required between two and three adaptive iterations depending on the geometry family and surrogate type, with most cases converging after three iterations. No systematic difference in the number of iterations was observed between RBF and supervised ML surrogates.
\begin{table*}[ht!]
\centering
......@@ -397,7 +399,7 @@ $F_5$ & Supervised ML & 3
\end{tabular}
\end{table*}
The final validation results show that both surrogate strategies provide FEM-consistent optimized geometries after only a few adaptive iterations. RBF surrogates generally lead to lower objective-function discrepancies, with an average $|e_J|$ of 1.36, compared with 5.04 for the supervised ML surrogates. The average maximum variable error is also lower for RBF, with 1.79\% compared with 2.67\% for supervised ML. The RBF surrogate provides particularly accurate predictions for the two-window devices, with $e_{\max}$ below 0.3\%, while remaining below 4.3\% for the three- and five-window families. The supervised ML surrogates also provide FEM-consistent results, although larger discrepancies are observed in some cases, especially for the $F_5$ family, where the objective-function error reaches 9.97.
The final validation results show that both surrogate strategies provide FEM-consistent optimized geometries after only a few adaptive iterations. For the set of optimized candidates analysed here, RBF surrogates generally lead to lower objective-function discrepancies, with an average $|e_J|$ of 1.36, compared with 5.04 for the supervised ML surrogates. The average maximum variable error is also lower for RBF, with 1.79\% compared with 2.67\% for supervised ML. The RBF surrogate provides particularly accurate predictions for the two-window devices, with $e_{\max}$ below 0.3\%, while remaining below 4.3\% for the three- and five-window families. The supervised ML surrogates also provide FEM-consistent results, although larger discrepancies are observed in some cases, especially for the $F_5$ family, where the objective-function error reaches 9.97. These figures quantify surrogate--FEM agreement at the optimized candidates and are therefore a local validation of the accepted designs; they are not a global ranking of predictive accuracy over the design domain, for which the outer cross-validation results discussed above should be consulted.
The adaptive validation loop is essential to reach these levels of agreement. In several cases, the first surrogate-optimized candidate did not satisfy the prescribed error tolerances, particularly for the three- and five-window devices. After incorporating the additional FEM results and retraining the surrogates, the prediction errors decreased significantly. For instance, the maximum variable error of the RBF surrogate decreased from 13.70\% to 1.70\% in the $F_3$ family and from 11.44\% to 2.66\% in the $F_5$ family. A similar behaviour was observed for the supervised ML surrogates, whose final candidates also reached a comparable level of agreement. This confirms that the adaptive loop reduces the risk of accepting geometries that appear optimal only because of surrogate prediction errors in sparsely sampled regions of the design space.
......@@ -412,11 +414,11 @@ Figure~\ref{fig:optimized_window_thickness_evolution} shows the evolution of the
\label{fig:optimized_window_thickness_evolution}
\end{figure*}
From a methodological point of view, these results highlight the trade-off between surrogate complexity, accuracy and computational efficiency. Supervised models, particularly SVR and GPR, provide high predictive accuracy and robustness, but require hyperparameter optimization and cross-validation for every output variable and adaptive iteration. RBF interpolation, in contrast, has a much lower training cost and provides very competitive final predictions once the relevant regions of the design domain have been adaptively sampled. Therefore, the comparison does not identify a universally superior surrogate strategy. Instead, it suggests that RBF interpolation is especially suitable for low- to moderate-dimensional design spaces with well-distributed FEM samples and relatively smooth input--output relationships, as occurs for the response variables analysed in this work. Supervised ML surrogates remain valuable when greater robustness is required or when the response surface is expected to involve stronger nonlinear interactions, local irregularities or higher-dimensional dependencies.
From a methodological point of view, these results highlight the trade-off between surrogate complexity, accuracy and computational efficiency. Supervised models, particularly SVR and GPR, provide competitive predictive accuracy under the adopted validation procedure, but require hyperparameter optimization and cross-validation for every output variable and adaptive iteration. RBF interpolation, in contrast, has a much lower training cost and provides very competitive final predictions once the relevant regions of the design domain have been adaptively sampled. Therefore, the comparison does not identify a universally superior surrogate strategy. Instead, it suggests that RBF interpolation is especially suitable for low- to moderate-dimensional design spaces with well-distributed FEM samples and relatively smooth input--output relationships, as occurs for the response variables analysed in this work. Supervised ML surrogates remain valuable when greater robustness is required or when the response surface is expected to involve stronger nonlinear interactions, local irregularities or higher-dimensional dependencies.
The objective-function values should also be interpreted carefully. The objective-function error measures the consistency between surrogate predictions and FEM validation, not whether the final objective value is necessarily close to zero. Some geometry families, such as $F_2$, retain non-negligible penalty contributions because the prescribed damage targets cannot be fully achieved within the admissible design bounds. Nevertheless, the close agreement between surrogate and FEM objective values indicates that the accepted designs are not artifacts of surrogate extrapolation, but FEM-consistent optimized candidates within the explored design space.
This behaviour is illustrated in Figure~\ref{fig:rbf_surface_evolution}, which shows the evolution of the RBF objective surface during the adaptive optimization process for the two-window families. The response surfaces remain relatively smooth, supporting the suitability of RBF interpolation for this problem. The evolution between adaptive iterations also shows how the surrogate surface is progressively corrected as new FEM information is incorporated near the optimized region.
This behaviour is illustrated in Figure~\ref{fig:rbf_surface_evolution}, which shows the evolution of the RBF objective surface during the adaptive optimization process for the two-window families. The response surfaces remain relatively smooth for the two-window families considered, consistently with the good performance of RBF interpolation observed for these datasets. The evolution between adaptive iterations also shows how the surrogate surface is progressively corrected as new FEM information is incorporated near the optimized region.
\begin{figure*}[htbp]
\centering
......@@ -431,7 +433,7 @@ This work presents an adaptive surrogate-assisted optimization framework for the
One of the key features of the proposed optimization framework, compared to approaches focused primarily on energy maximization, is the formulation of an objective function that takes damage into account. The aim is to prioritise local damage control in both the dissipative windows and the surrounding frame, whilst promoting a balanced contribution from all windows to the energy dissipation process. Thus, damage to the windows is permitted and expected, as they are intended to act as dissipative regions, provided that it remains controlled and reasonably distributed; conversely, damage to the frame is penalised more severely because it can compromise the structural integrity of the damper. Therefore, rather than merely maximizing the energy dissipated, the proposed formulation favours geometries that concentrate dissipative activation in the windows, prevent excessive damage localization in a single region and protect the frame from critical damage.
A comprehensive comparison of surrogate strategies was performed in terms of predictive accuracy and computational efficiency. The supervised ML results showed that SVR and GPR were the most competitive models, with SVR being the most frequently selected across the analysed outputs. RBF interpolation proved to be even a better alternative, due to its high efficiency: in the final adaptive iterations, it achieved validation errors comparable to, and in most cases lower than, those of the supervised ML surrogates, while requiring substantially lower training effort and no hyperparameter search. This performance is attributed to the characteristics of the present problem: low- to moderate-dimensional design spaces, well-distributed FEM samples and nonlinear but relatively smooth relationships between window thicknesses and response indicators.
A comparison of surrogate strategies was performed in terms of predictive accuracy and computational efficiency. For the analysed datasets, the supervised ML results showed that kernel-based models, particularly SVR and GPR, were the most frequently selected, with SVR being the most frequently selected across the outputs. When evaluated on the same outer cross-validation splits, RBF interpolation provided predictive performance comparable to the supervised models for the considered low-dimensional response surfaces, while requiring a much smaller hyperparameter search and substantially lower training effort. For the optimized candidates, the RBF surrogate also achieved surrogate--FEM agreement comparable to, and in several cases better than, that of the supervised ML surrogates. These observations are limited to the analysed geometry families and to the sampled design domain, and their reliability is lower for the smallest datasets, for which the accuracy estimates carry larger uncertainty.
The proposed adaptive validation loop proved to be necessary and effective. Several initially optimized candidates did not satisfy the prescribed error tolerances. After incorporating the new FEM results into the training dataset and retraining the surrogates, the prediction errors decreased and the optimization converged after only two or three iterations. Therefore, the final designs are not accepted solely on the basis of surrogate predictions, but are explicitly verified through FEM in the region of the design space where the optimum is located.
......
ManuscriptR1/ComparisonSurrogatesOptimizationBDSL_R1_changes.tex
\ifdefined\pdfminorversion
%DIF LATEXDIFF DIFFERENCE FILE
%DIF DEL ComparisonSurrogatesOptimizationBDSL_submitted.tex Mon Oct 5 15:02:51 2026
%DIF ADD ComparisonSurrogatesOptimizationBDSL_R1.tex Mon Oct 5 16:38:21 2026
%DIF ADD ComparisonSurrogatesOptimizationBDSL_R1.tex Mon Oct 5 16:49:41 2026
\pdfminorversion=7
\fi
\documentclass[final,5p,times,twocolumn]{elsarticle}
......@@ -108,12 +108,12 @@
country={Spain}}
\begin{abstract}
Buckling-delayed shear-link dampers are passive devices used in seismic-resistant structures to concentrate energy dissipation while limiting damage to the primary system. \DIFdelbegin \DIFdel{Their geometric optimization }\DIFdelend \DIFaddbegin \DIFadd{Optimizing their window thicknesses }\DIFaddend requires balancing high dissipative capacity with strict control of local damage. Nonlinear finite element models can reproduce their cyclic response and provide internal quantities for optimization, such as damage indicators and local distortion, but their computational cost prevents direct use within iterative optimization loops. This work proposes an adaptive surrogate-assisted optimization framework for \DIFaddbegin \DIFadd{damage-aware window-thickness optimization of }\DIFaddend buckling-delayed shear-link dampers \DIFaddbegin \DIFadd{within five predefined geometry families}\DIFaddend . Experimentally calibrated nonlinear finite element models are first used to generate reference datasets for different device configurations. Supervised learning models are evaluated, with support vector regression and Gaussian process regression showing high predictive accuracy, motivating the use of radial basis function surrogates as a more efficient distance-based alternative. The surrogate predictions are coupled with a differential evolution algorithm through a damage-aware objective function that controls local damage while using dissipative performance as an optimization criterion. The optimized geometries are then re-evaluated with finite element simulations. If the \DIFdelbegin \DIFdel{surrogate error exceeds the adopted tolerances}\DIFdelend \DIFaddbegin \DIFadd{acceptance criteria are not met}\DIFaddend , the new simulation is added to the dataset and the surrogate models are retrained. The framework enables efficient damage-aware optimization of seismic energy dissipation devices.
Buckling-delayed shear-link dampers are passive devices used in seismic-resistant structures to concentrate energy dissipation while limiting damage to the primary system. \DIFdelbegin \DIFdel{Their geometric optimization }\DIFdelend \DIFaddbegin \DIFadd{Optimizing their window thicknesses }\DIFaddend requires balancing high dissipative capacity with strict control of local damage. Nonlinear finite element models can reproduce their cyclic response and provide internal quantities for optimization, such as damage indicators and local distortion, but their computational cost prevents direct use within iterative optimization loops. This work proposes an adaptive surrogate-assisted optimization framework for \DIFaddbegin \DIFadd{damage-aware window-thickness optimization of }\DIFaddend buckling-delayed shear-link dampers \DIFaddbegin \DIFadd{within five predefined geometry families}\DIFaddend . Experimentally calibrated nonlinear finite element models are first used to generate reference datasets for different device configurations. Supervised learning models are evaluated, with support vector regression and Gaussian process regression \DIFdelbegin \DIFdel{showing high predictive accuracy}\DIFdelend \DIFaddbegin \DIFadd{providing the most accurate predictions among the candidate models}\DIFaddend , motivating the use of radial basis function surrogates as a more efficient distance-based alternative. The surrogate predictions are coupled with a differential evolution algorithm through a damage-aware objective function that controls local damage while using dissipative performance as an optimization criterion. The optimized geometries are then re-evaluated with finite element simulations. If the \DIFdelbegin \DIFdel{surrogate error exceeds the adopted tolerances}\DIFdelend \DIFaddbegin \DIFadd{acceptance criteria are not met}\DIFaddend , the new simulation is added to the dataset and the surrogate models are retrained. The framework enables efficient damage-aware optimization of seismic energy dissipation devices.
\end{abstract}
\begin{highlights}
\item Adaptive surrogates optimize damage-aware shear-link damper geometries
\item RBF interpolation matches supervised ML with lower retraining cost
\item RBF interpolation \DIFdelbegin \DIFdel{matches supervised ML with lower retraining }\DIFdelend \DIFaddbegin \DIFadd{offers comparable accuracy to supervised ML at lower training }\DIFaddend cost
\item FEM validation filters surrogate optima before accepting final designs
\item Optimized devices balance window activation and frame damage control
\end{highlights}
......@@ -374,7 +374,7 @@ The surrogate-optimized geometry is not accepted directly. Instead, once an opti
\section{Numerical results and discussion}\label{sec:results}
The supervised-learning comparison shows a clear hierarchy among the candidate surrogate models. A total of 100 output-specific training problems were considered, corresponding to the $2N_w+1$ target variables required for each geometry family and adaptive iteration. The $F_1$ family required two optimization iterations, whereas the remaining families required three. Across all geometry families, iterations and output variables, SVR was the most frequently selected model and also the model that most often achieved the lowest cross-validated RMSE. As summarized in Figure~\ref{fig:surrogate_selection_summary_barplot}, SVR was selected in 71 cases, followed by GPR, GBR, XGBoost and MLP, while Random Forest was not selected in any case. This dominance was particularly clear for the two-window families, where only one output was assigned to a GPR model and for the damage-related outputs, for which SVR was selected in 47 out of 57 cases. For more complex devices, those with three and five windows, and for distortion-related outputs, the model selection became more heterogeneous, although SVR still provided the best overall performance. From the computational point of view, SVR also offered the lowest median training times among the supervised models, whereas MLP required substantially longer training times without providing gain in accuracy.
The supervised-learning comparison shows a \DIFdelbegin \DIFdel{clear hierarchy }\DIFdelend \DIFaddbegin \DIFadd{predominance of kernel-based models }\DIFaddend among the candidate \DIFdelbegin \DIFdel{surrogate models}\DIFdelend \DIFaddbegin \DIFadd{surrogates}\DIFaddend . A total of 100 output-specific \DIFdelbegin \DIFdel{training }\DIFdelend \DIFaddbegin \DIFadd{model-selection }\DIFaddend problems were considered, corresponding to the $2N_w+1$ target variables required for each geometry family and adaptive iteration. The $F_1$ family required two optimization iterations, whereas the remaining families required three. Across all geometry families, iterations and output variables, SVR was the most frequently selected model and also the model that most often achieved the lowest cross-validated RMSE. As summarized in Figure~\ref{fig:surrogate_selection_summary_barplot}, SVR was selected in 71 cases, followed by GPR \DIFdelbegin \DIFdel{, GBR, XGBoost and MLP}\DIFdelend \DIFaddbegin \DIFadd{(15), GBR (10), XGBoost (2) and MLP (2)}\DIFaddend , while Random Forest was not selected in any case. This \DIFdelbegin \DIFdel{dominance }\DIFdelend \DIFaddbegin \DIFadd{predominance }\DIFaddend was particularly clear for the two-window families, where only one output was assigned to a GPR model\DIFaddbegin \DIFadd{, }\DIFaddend and for the damage-related outputs, for which SVR was selected in 47 out of 57 cases. For more complex devices, those with three and five windows, and for distortion-related outputs, the model selection became more heterogeneous, although SVR \DIFdelbegin \DIFdel{still provided the best overall performance}\DIFdelend \DIFaddbegin \DIFadd{remained the most frequently selected model. The RMSE-dispersion selection criterion caused the selected model to differ from the single lowest-RMSE model only in a minority of the aggregate counts; for example, GPR attained the lowest RMSE in 17 outputs but was selected in 15, whereas GBR was selected in 10 and attained the lowest RMSE in 9. These counts are descriptive statistics over model-selection tasks that are not mutually independent, because several outputs belong to the same geometry family, adaptive iteration and correlated FEM response. They are therefore reported as a descriptive summary of the comparison rather than as independent statistical evidence of superiority, and the sensitivity of the selection to the adopted 5\% competitive RMSE threshold has not been assessed}\DIFaddend . From the computational point of view, SVR also offered the lowest median training times among the supervised models, whereas MLP required substantially longer training times without providing gain in accuracy.
\begin{figure*}[htbp]
\centering
......@@ -383,9 +383,12 @@ The supervised-learning comparison shows a clear hierarchy among the candidate s
\label{fig:surrogate_selection_summary_barplot}
\end{figure*}
These results indicate that kernel-based models are particularly well suited to the present surrogate task. SVR provides the best compromise between accuracy and computational cost, while GPR is the second most competitive supervised strategy, especially in some higher-dimensional cases. Tree-based models, although robust, are less frequently selected and MLP models are not competitive in terms of computational efficiency for the dataset sizes considered here. As mentioned in previous sections, this behaviour motivated the additional evaluation of RBF interpolation as a simpler surrogate alternative. In contrast to the supervised models, RBF models were trained in less than one second per output, making them especially attractive for repeated surrogate updates within the adaptive optimization loop.
These results indicate that kernel-based models \DIFdelbegin \DIFdel{are particularly well suited to the present surrogate task. SVR provides }\DIFdelend \DIFaddbegin \DIFadd{performed well under the adopted nested validation procedure for the datasets considered here. SVR provided }\DIFaddend the best compromise between accuracy and computational cost, while GPR \DIFdelbegin \DIFdel{is }\DIFdelend \DIFaddbegin \DIFadd{was }\DIFaddend the second most \DIFdelbegin \DIFdel{competitive }\DIFdelend \DIFaddbegin \DIFadd{frequently selected }\DIFaddend supervised strategy, especially in some higher-dimensional cases. Tree-based models \DIFdelbegin \DIFdel{, although robust, are }\DIFdelend \DIFaddbegin \DIFadd{were }\DIFaddend less frequently selected and MLP models \DIFdelbegin \DIFdel{are }\DIFdelend \DIFaddbegin \DIFadd{were }\DIFaddend not competitive in terms of computational efficiency for the dataset sizes considered\DIFdelbegin \DIFdel{here}\DIFdelend . As mentioned in previous sections, this behaviour motivated the additional evaluation of RBF interpolation as a simpler surrogate alternative. \DIFdelbegin \DIFdel{In contrast to }\DIFdelend \DIFaddbegin \DIFadd{Because the RBF and }\DIFaddend the supervised models \DIFdelbegin \DIFdel{, RBF models were trained in less than one second per output , making them especially }\DIFdelend \DIFaddbegin \DIFadd{are assessed on the same outer cross-validation splits and with the same error metrics, the two strategies can be compared on a like-for-like basis at the level of each output variable. Across the analysed outputs neither strategy dominates: the selected supervised model is more accurate for some targets and the RBF surrogate for others, and both remain within the accuracy range required by the adaptive validation loop. The RBF surrogate, however, is obtained with a much smaller hyperparameter search and therefore requires substantially lower training effort, which makes it }\DIFaddend attractive for repeated surrogate updates within the adaptive optimization loop.
The FEM validation of the optimized geometries is summarized in Table~\ref{tab:final_surrogate_comparison}, while the complete optimization results for all adaptive iterations are provided in Appendix~\ref{appendix:optimization_results}. For each geometry family and surrogate strategy, Table~\ref{tab:final_surrogate_comparison} reports the final accepted adaptive iteration, the optimized window thicknesses $\mathbf{t}_w^{\star}$, the surrogate-predicted objective value $J_{\mathrm{surr}}$, the corresponding FEM-recomputed objective value $J_{\mathrm{FEM}}$ and the associated validation errors, $|e_J|$ and $e_{\max}$. The maximum variable error, $e_{\max}$, is defined as the largest relative error among all quantities entering the objective function, namely ${\Exy}_i$, $\TFD_i$ and $\TFD_f$, whereas $|e_J|$ denotes the absolute objective-function error. The optimization required between two and three adaptive iterations depending on the geometry family and surrogate type, with most cases converging after three iterations. No systematic difference in the number of iterations was observed between RBF and supervised ML surrogates.
The \DIFaddbegin \DIFadd{predictive assessment is constrained by the small datasets available for the two-window families ($N=8$) and the three-window families ($N=16$). Nested cross-validation separates hyperparameter selection from generalization assessment, and the outer folds provide out-of-sample predictions together with bootstrap confidence intervals for RMSE and MAE. For datasets with $N\leq20$ the outer loop is Leave-One-Out, so the reported errors are obtained from a single deterministic partition and their dispersion is estimated by resampling the out-of-sample predictions rather than by repeating the outer split. These results should therefore be interpreted as an assessment of predictive accuracy within the sampled design domain and for the analysed geometry families, under the adopted validation procedure, rather than as a general guarantee. No independent FEM test set was retained, and the uncertainty of the accuracy estimates for the smallest datasets remains a limitation of the present comparison.
}
\DIFadd{The }\DIFaddend FEM validation of the optimized geometries is summarized in Table~\ref{tab:final_surrogate_comparison}, while the complete optimization results for all adaptive iterations are provided in Appendix~\ref{appendix:optimization_results}. For each geometry family and surrogate strategy, Table~\ref{tab:final_surrogate_comparison} reports the final accepted adaptive iteration, the optimized window thicknesses $\mathbf{t}_w^{\star}$, the surrogate-predicted objective value $J_{\mathrm{surr}}$, the corresponding FEM-recomputed objective value $J_{\mathrm{FEM}}$ and the associated validation errors, $|e_J|$ and $e_{\max}$. The maximum variable error, $e_{\max}$, is defined as the largest relative error among all quantities entering the objective function, namely ${\Exy}_i$, $\TFD_i$ and $\TFD_f$, whereas $|e_J|$ denotes the absolute objective-function error. \DIFaddbegin \DIFadd{It should be emphasised that these errors measure the agreement between the surrogate and the FEM model at the optimized candidate; they characterise the local consistency of the surrogate in the vicinity of the optimum and are not, by themselves, a measure of global predictive accuracy over the whole design domain. }\DIFaddend The optimization required between two and three adaptive iterations depending on the geometry family and surrogate type, with most cases converging after three iterations. No systematic difference in the number of iterations was observed between RBF and supervised ML surrogates.
\begin{table*}[ht!]
\centering
......@@ -439,7 +442,8 @@ $F_5$ & Supervised ML & 3
\end{tabular}
\end{table*}
The final validation results show that both surrogate strategies provide FEM-consistent optimized geometries after only a few adaptive iterations. RBF surrogates generally lead to lower objective-function discrepancies, with an average $|e_J|$ of 1.36, compared with 5.04 for the supervised ML surrogates. The average maximum variable error is also lower for RBF, with 1.79\% compared with 2.67\% for supervised ML. The RBF surrogate provides particularly accurate predictions for the two-window devices, with $e_{\max}$ below 0.3\%, while remaining below 4.3\% for the three- and five-window families. The supervised ML surrogates also \DIFdelbegin \DIFdel{satisfy all validation criteria}\DIFdelend \DIFaddbegin \DIFadd{provide FEM-consistent results}\DIFaddend , although larger discrepancies are observed in some cases, especially for the $F_5$ family, where the objective-function error reaches 9.97.
The final validation results show that both surrogate strategies provide FEM-consistent optimized geometries after only a few adaptive iterations. \DIFaddbegin \DIFadd{For the set of optimized candidates analysed here, }\DIFaddend RBF surrogates generally lead to lower objective-function discrepancies, with an average $|e_J|$ of 1.36, compared with 5.04 for the supervised ML surrogates. The average maximum variable error is also lower for RBF, with 1.79\% compared with 2.67\% for supervised ML. The RBF surrogate provides particularly accurate predictions for the two-window devices, with $e_{\max}$ below 0.3\%, while remaining below 4.3\% for the three- and five-window families. The supervised ML surrogates also \DIFdelbegin \DIFdel{satisfy all validation criteria}\DIFdelend \DIFaddbegin \DIFadd{provide FEM-consistent results}\DIFaddend , although larger discrepancies are observed in some cases, especially for the $F_5$ family, where the objective-function error reaches 9.97. \DIFaddbegin \DIFadd{These figures quantify surrogate--FEM agreement at the optimized candidates and are therefore a local validation of the accepted designs; they are not a global ranking of predictive accuracy over the design domain, for which the outer cross-validation results discussed above should be consulted.
}\DIFaddend
The adaptive validation loop is essential to reach these levels of agreement. In several cases, the first surrogate-optimized candidate did not satisfy the prescribed error tolerances, particularly for the three- and five-window devices. After incorporating the additional FEM results and retraining the surrogates, the prediction errors decreased significantly. For instance, the maximum variable error of the RBF surrogate decreased from 13.70\% to 1.70\% in the $F_3$ family and from 11.44\% to 2.66\% in the $F_5$ family. A similar behaviour was observed for the supervised ML surrogates, whose final candidates also \DIFdelbegin \DIFdel{satisfied all acceptance criteria}\DIFdelend \DIFaddbegin \DIFadd{reached a comparable level of agreement}\DIFaddend . This confirms that the adaptive loop reduces the risk of accepting geometries that appear optimal only because of surrogate prediction errors in sparsely sampled regions of the design space.
......@@ -454,11 +458,11 @@ Figure~\ref{fig:optimized_window_thickness_evolution} shows the evolution of the
\label{fig:optimized_window_thickness_evolution}
\end{figure*}
From a methodological point of view, these results highlight the trade-off between surrogate complexity, accuracy and computational efficiency. Supervised models, particularly SVR and GPR, provide high predictive accuracy and robustness, but require hyperparameter optimization and cross-validation for every output variable and adaptive iteration. RBF interpolation, in contrast, has a much lower training cost and provides very competitive final predictions once the relevant regions of the design domain have been adaptively sampled. Therefore, the comparison does not identify a universally superior surrogate strategy. Instead, it suggests that RBF interpolation is especially suitable for low- to moderate-dimensional design spaces with well-distributed FEM samples and relatively smooth input--output relationships, as occurs for the response variables analysed in this work. Supervised ML surrogates remain valuable when greater robustness is required or when the response surface is expected to involve stronger nonlinear interactions, local irregularities or higher-dimensional dependencies.
From a methodological point of view, these results highlight the trade-off between surrogate complexity, accuracy and computational efficiency. Supervised models, particularly SVR and GPR, provide \DIFdelbegin \DIFdel{high predictive accuracy and robustness}\DIFdelend \DIFaddbegin \DIFadd{competitive predictive accuracy under the adopted validation procedure}\DIFaddend , but require hyperparameter optimization and cross-validation for every output variable and adaptive iteration. RBF interpolation, in contrast, has a much lower training cost and provides very competitive final predictions once the relevant regions of the design domain have been adaptively sampled. Therefore, the comparison does not identify a universally superior surrogate strategy. Instead, it suggests that RBF interpolation is especially suitable for low- to moderate-dimensional design spaces with well-distributed FEM samples and relatively smooth input--output relationships, as occurs for the response variables analysed in this work. Supervised ML surrogates remain valuable when greater robustness is required or when the response surface is expected to involve stronger nonlinear interactions, local irregularities or higher-dimensional dependencies.
The objective-function values should also be interpreted carefully. The objective-function error measures the consistency between surrogate predictions and FEM validation, not whether the final objective value is necessarily close to zero. Some geometry families, such as $F_2$, retain non-negligible penalty contributions because the prescribed damage targets cannot be fully achieved within the admissible design bounds. Nevertheless, the close agreement between surrogate and FEM objective values indicates that the accepted designs are not artifacts of surrogate extrapolation, but FEM-consistent optimized candidates within the explored design space.
This behaviour is illustrated in Figure~\ref{fig:rbf_surface_evolution}, which shows the evolution of the RBF objective surface during the adaptive optimization process for the two-window families. The response surfaces remain relatively smooth, supporting the suitability of RBF interpolation for this problem. The evolution between adaptive iterations also shows how the surrogate surface is progressively corrected as new FEM information is incorporated near the optimized region.
This behaviour is illustrated in Figure~\ref{fig:rbf_surface_evolution}, which shows the evolution of the RBF objective surface during the adaptive optimization process for the two-window families. The response surfaces remain relatively smooth \DIFdelbegin \DIFdel{, supporting the suitability }\DIFdelend \DIFaddbegin \DIFadd{for the two-window families considered, consistently with the good performance }\DIFaddend of RBF interpolation \DIFdelbegin \DIFdel{for this problem}\DIFdelend \DIFaddbegin \DIFadd{observed for these datasets}\DIFaddend . The evolution between adaptive iterations also shows how the surrogate surface is progressively corrected as new FEM information is incorporated near the optimized region.
\begin{figure*}[htbp]
\centering
......@@ -473,7 +477,7 @@ This work presents an adaptive surrogate-assisted optimization framework for \DI
One of the key features of the proposed optimization framework, compared to approaches focused primarily on energy maximization, is the formulation of an objective function that takes damage into account. The aim is to prioritise local damage control in both the dissipative windows and the surrounding frame, whilst promoting a balanced contribution from all windows to the energy dissipation process. Thus, damage to the windows is permitted and expected, as they are intended to act as dissipative regions, provided that it remains controlled and reasonably distributed; conversely, damage to the frame is penalised more severely because it can compromise the structural integrity of the damper. Therefore, rather than merely maximizing the energy dissipated, the proposed formulation favours geometries that concentrate dissipative activation in the windows, prevent excessive damage localization in a single region and protect the frame from critical damage.
A comprehensive comparison of surrogate strategies was performed in terms of predictive accuracy and computational efficiency. The supervised ML results showed that SVR and GPR were the most competitive models, with SVR being the most frequently selected across the analysed outputs. RBF interpolation proved to be even a better alternative, due to its high efficiency: in the final adaptive iterations, it achieved validation errors comparable to, and in most cases lower than, those of the supervised ML surrogates, while requiring substantially lower training effort and no hyperparameter search. This performance is attributed to the characteristics of the present problem: low- to moderate-dimensional design spaces, well-distributed FEM samples and nonlinear but relatively smooth relationships between window thicknesses and response indicators.
A \DIFdelbegin \DIFdel{comprehensive }\DIFdelend comparison of surrogate strategies was performed in terms of predictive accuracy and computational efficiency. \DIFdelbegin \DIFdel{The }\DIFdelend \DIFaddbegin \DIFadd{For the analysed datasets, the }\DIFaddend supervised ML results showed that \DIFaddbegin \DIFadd{kernel-based models, particularly }\DIFaddend SVR and GPR\DIFaddbegin \DIFadd{, }\DIFaddend were the most \DIFdelbegin \DIFdel{competitive models}\DIFdelend \DIFaddbegin \DIFadd{frequently selected}\DIFaddend , with SVR being the most frequently selected across the \DIFdelbegin \DIFdel{analysed outputs. RBF interpolation proved to be even a better alternative, due to its high efficiency: in the final adaptive iterations, it achieved validation errors }\DIFdelend \DIFaddbegin \DIFadd{outputs. When evaluated on the same outer cross-validation splits, RBF interpolation provided predictive performance comparable to the supervised models for the considered low-dimensional response surfaces, while requiring a much smaller hyperparameter search and substantially lower training effort. For the optimized candidates, the RBF surrogate also achieved surrogate--FEM agreement }\DIFaddend comparable to, and in \DIFdelbegin \DIFdel{most cases lower than, those }\DIFdelend \DIFaddbegin \DIFadd{several cases better than, that }\DIFaddend of the supervised ML surrogates\DIFdelbegin \DIFdel{, while requiring substantially lower training effort and no hyperparameter search. This performance is attributed to the characteristics of the present problem: low- to moderate-dimensional design spaces, well-distributed FEM samples and nonlinear but relatively smooth relationships between window thicknesses and response indicators}\DIFdelend \DIFaddbegin \DIFadd{. These observations are limited to the analysed geometry families and to the sampled design domain, and their reliability is lower for the smallest datasets, for which the accuracy estimates carry larger uncertainty}\DIFaddend .
The proposed adaptive validation loop proved to be necessary and effective. Several initially optimized candidates did not satisfy the prescribed error tolerances. After incorporating the new FEM results into the training dataset and retraining the surrogates, the prediction errors decreased and \DIFdelbegin \DIFdel{all final optimized geometries satisfied the acceptance criteria }\DIFdelend \DIFaddbegin \DIFadd{the optimization converged }\DIFaddend after only two or three iterations. Therefore, the final designs are not accepted solely on the basis of surrogate predictions, but are explicitly verified through FEM in the region of the design space where the optimum is located.
......
ManuscriptR1/images/BayesianSearchCV/BayesianSearchCV.drawio
......@@ -692,7 +692,7 @@
<mxCell id="ulNCGQqmzGW29SgBMLZ--3" parent="1" style="text;html=1;whiteSpace=wrap;strokeColor=none;fillColor=none;align=center;verticalAlign=middle;rounded=0;fontColor=light-dark(#FF8000,#FF8000);" value="&lt;font style=&quot;font-size: 22px;&quot;&gt;&lt;b&gt;&lt;font style=&quot;&quot;&gt;If 21 ≤&amp;nbsp;N&amp;nbsp;≤ 8&lt;/font&gt;&lt;span style=&quot;background-color: transparent;&quot;&gt;0&lt;/span&gt;&lt;/b&gt;&lt;/font&gt;" vertex="1">
<mxGeometry height="30" width="140" x="-145" y="1048" as="geometry" />
</mxCell>
<mxCell id="ulNCGQqmzGW29SgBMLZ--4" parent="1" style="text;html=1;whiteSpace=wrap;strokeColor=none;fillColor=none;align=center;verticalAlign=middle;rounded=0;fontSize=18;fontStyle=1;fontColor=light-dark(#000000,#000000);" value="&lt;font&gt;Repeated K-Fold CV (&lt;i&gt;K&amp;nbsp;&lt;/i&gt;= 4, 5 times)&lt;/font&gt;" vertex="1">
<mxCell id="ulNCGQqmzGW29SgBMLZ--4" parent="1" style="text;html=1;whiteSpace=wrap;strokeColor=none;fillColor=none;align=center;verticalAlign=middle;rounded=0;fontSize=18;fontStyle=1;fontColor=light-dark(#000000,#000000);" value="&lt;font&gt;Repeated K-Fold CV (&lt;i&gt;K&amp;nbsp;&lt;/i&gt;= 5, 3 times)&lt;/font&gt;" vertex="1">
<mxGeometry height="44" width="200" x="-175" y="1081" as="geometry" />
</mxCell>
<mxCell id="ulNCGQqmzGW29SgBMLZ--5" parent="1" style="text;html=1;whiteSpace=wrap;strokeColor=none;fillColor=none;align=center;verticalAlign=middle;rounded=0;fontColor=light-dark(#009900,#009900);" value="&lt;font style=&quot;font-size: 22px;&quot;&gt;&lt;b&gt;&lt;font style=&quot;&quot;&gt;If N &amp;gt; 8&lt;/font&gt;&lt;span style=&quot;background-color: transparent;&quot;&gt;0&lt;/span&gt;&lt;/b&gt;&lt;/font&gt;" vertex="1">
......@@ -719,7 +719,7 @@
<mxCell id="ulNCGQqmzGW29SgBMLZ--14" parent="1" style="text;html=1;whiteSpace=wrap;strokeColor=none;fillColor=none;align=left;verticalAlign=middle;rounded=0;fontColor=light-dark(#000000,#000000);fontSize=20;" value="&lt;font&gt;Multi-Layer Perceptron (MLP)&lt;/font&gt;" vertex="1">
<mxGeometry height="50" width="180" x="340" y="1171" as="geometry" />
</mxCell>
<mxCell id="ulNCGQqmzGW29SgBMLZ--15" parent="1" style="text;html=1;whiteSpace=wrap;strokeColor=none;fillColor=none;align=center;verticalAlign=middle;rounded=0;fontSize=20;" value="&lt;div&gt;&lt;font style=&quot;color: light-dark(rgb(0, 0, 0), rgb(0, 0, 0));&quot;&gt;(40 evaluations)&lt;/font&gt;&lt;/div&gt;&lt;div&gt;&lt;font style=&quot;color: light-dark(rgb(0, 0, 0), rgb(0, 0, 0));&quot;&gt;&lt;br&gt;&lt;/font&gt;&lt;/div&gt;&lt;div&gt;&lt;font&gt;&lt;font style=&quot;background-color: transparent; color: light-dark(rgb(0, 0, 0), rgb(0, 0, 0));&quot;&gt;Cross-validated&amp;nbsp;&lt;/font&gt;&lt;span style=&quot;background-color: transparent; color: light-dark(rgb(0, 0, 0), rgb(0, 0, 0));&quot;&gt;RMSE for each configuration&lt;/span&gt;&lt;/font&gt;&lt;/div&gt;" vertex="1">
<mxCell id="ulNCGQqmzGW29SgBMLZ--15" parent="1" style="text;html=1;whiteSpace=wrap;strokeColor=none;fillColor=none;align=center;verticalAlign=middle;rounded=0;fontSize=20;" value="&lt;div&gt;&lt;font style=&quot;color: light-dark(rgb(0, 0, 0), rgb(0, 0, 0));&quot;&gt;(20, 30 or 40 evaluations)&lt;/font&gt;&lt;/div&gt;&lt;div&gt;&lt;font style=&quot;color: light-dark(rgb(0, 0, 0), rgb(0, 0, 0));&quot;&gt;&lt;br&gt;&lt;/font&gt;&lt;/div&gt;&lt;div&gt;&lt;font&gt;&lt;font style=&quot;background-color: transparent; color: light-dark(rgb(0, 0, 0), rgb(0, 0, 0));&quot;&gt;Cross-validated&amp;nbsp;&lt;/font&gt;&lt;span style=&quot;background-color: transparent; color: light-dark(rgb(0, 0, 0), rgb(0, 0, 0));&quot;&gt;RMSE for each configuration&lt;/span&gt;&lt;/font&gt;&lt;/div&gt;" vertex="1">
<mxGeometry height="110" width="212" x="554" y="791" as="geometry" />
</mxCell>
<mxCell id="ulNCGQqmzGW29SgBMLZ--16" parent="1" style="shape=image;aspect=fixed;image=data:image/png,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;" value="" vertex="1">
......
ManuscriptR1/images/BayesianSearchCV/BayesianSearchCV.svg
This source diff could not be displayed because it is too large. You can view the blob instead.
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment