Chani-Song/wafer-jmd-release-data

WAFER battery-pack redesign runs

This dataset holds the result tables and the raw run records of "LLM-Assisted Optimization for Simulation-Based Redesign" by Chan-Eui Song, Jeremy J. Michalek and Christopher McComb, submitted to the ASME Journal of Mechanical Design (2026). The 12,400 runs redesign an electric vehicle battery pack with the BatPaC 5.0 simulator, using a language model for warm starting (WS), variable fixing (VF) or ReAct, or a numerical optimizer alone. The code is at https://github.com/Chani-song/WAFER.

Every table is in data/ as Parquet and as CSV, gzipped for evaluations. Each table loads by name:

import pandas as pd
runs = pd.read_parquet("hf://datasets/Chani-Song/wafer-jmd-release-data/data/runs.parquet")

Tables

Group	Table	Rows	One row is
Results	runs	12,400	one primary run of 20 evaluations, with its outcome
	configurations	248	one configuration of 50 runs, with its summary
	evaluations	239,645	one accepted BatPaC evaluation of a run, with all outputs and the wall time
Language model decisions	vf_decisions	4,800	the fixing decision of one VF run
	ws_candidates	14,331	one of the three starting designs proposed in a WS run
	react_steps	22,104	one accepted ReAct step with its reasoning text
	vf_rationale	4,800	the rationale text of one VF decision with its reasoning labels
	ws_rationale	19,131	one WS rationale text with its reasoning labels
	react_calls	1,200	one ReAct run with its call and token counts
	react_call_events	43,004	one event of a ReAct trace
Supporting	grid_evaluations	5,496	one evaluation of the preliminary design grid
	l2_comprehension_responses	12	one response of the L2 comprehension check

run_id links the tables. Every configuration has 50 runs, and every run had a budget of 20 evaluations. 204 runs ended before their first evaluation because the language model gave no valid decision, so evaluations covers 12,196 runs. In grid_evaluations, S2 feasibility is not stored, because S2 changes only the width limit: Table 7 recomputes it from the source rows with a width of at most 1,100 mm. The L2 comprehension responses come from a 4-bit MLX build of Qwen3.5-4B at temperature 0.7, top-p 0.8 and at most 3,500 tokens.

Codes

• Scenarios S1 to S4 change the requirements of the source design. S1 raises the NMC811 cathode material cost from 19.63 to 25.75 USD/kg. S2 lowers the maximum pack width from 1500 to 1100 mm. S3 shortens the recharge time from 60 to 30 min. S4 raises the required energy from 84 to 130 kWh and allows a pack of up to 2700 by 1650 mm. The experiment code calls them s_cost, s_space, s_fastcharge and s_energy130_relaxed_pack.
• Methods (strategy): Cold is the cold start, SW the source-warm start, VF variable fixing, WS warm starting and ReAct the ReAct agent.
• Optimizers: axbo is Bayesian optimization (BO), mixed_ga the genetic algorithm (GA), mads mesh adaptive direct search (MADS) and random random search. ReAct uses no optimizer.
• The source design is NMC811, 155 µm, 29 cells per module, 2 modules per row and 4 rows.
• A VF mask has five letters for x^Chem, x^T, x^S, x^MPR and x^R in that order. R marks a variable released to the optimizer and F one fixed at its source value, so RFFFF releases only the chemistry.
• semantic_level L1 gives the model the problem definition, and L2 adds engineering relationships. model_size is 4, 9 or 27 billion parameters of Qwen3.5.
• Feasible discovery is feasible_found and near-best attainment is within2pct. The reference cost is the lowest feasible cost over all primary evaluations of a scenario: 97.104 USD/kWh in S1, 97.398 in S2, 100.044 in S3 and 90.283 in S4.

Shared columns

Column	Type	Meaning
run_id	string	run identifier, the path of the run folder in the raw records. It links the tables.
scenario	string	S1, S2, S3 or S4
strategy	string	Cold, SW, VF, WS or ReAct
optimizer	string	axbo, mixed_ga, mads or random. Empty for ReAct.
model_size	float or int	4, 9 or 27. Empty for Cold and SW.
semantic_level	string	L1 or L2. Empty for Cold and SW.
seed	int	run index 0 to 49. The generation seed equals it, and the optimizer seed is 2000 plus it.
x_chem	string	cathode chemistry, LFP or NMC811
x_T	float, µm	cathode thickness, 50 to 165 µm for LFP and 50 to 155 µm for NMC811
x_S	int	cells per module (BatPaC Number of cells per module), 1 to 30
x_MPR	int	modules per row, 1 to 50
x_R	int	module rows, 1 to 4

Columns

runs

Shared columns: run_id, scenario, strategy, optimizer, model_size, semantic_level, seed.

Column	Type	Unit	Meaning	Values
feasible_found	int64		1 if any accepted evaluation was feasible (feasible discovery). Failed runs count as 0.	0, 1
first_feasible_eval	float64		index k of the first feasible accepted evaluation. Empty if none.	1 to 20
final_best_feasible_cost	float64	USD/kWh	lowest feasible cost of the run. Empty if none.	90.2828 to 145.409
within2pct	int64		1 if the run ended within 2% of the reference cost of its scenario (near-best attainment)	0, 1
within5pct	int64		1 if the run ended within 5% of the reference cost	0, 1
n_accepted_evals	int64		number of accepted BatPaC evaluations, at most 20	0 to 20
tokens_total	float64	tokens	prompt and completion tokens of the run as reported by the vLLM server. 0 for Cold and SW.	0 to 1.16402e+06
inference_time_total	float64	s	measured language model inference time of the run, without simulator time. 0 for Cold and SW.	0 to 2007.05
terminal_failure	int64		1 if the run ended without a usable decision (terminal_status failed)	0, 1
failure_reason	string		error that ended a failed run, verbatim	9 distinct values
protocol_valid	int64		1 if the run followed the protocol to its end	0, 1
source_design_cost	float64	USD/kWh	S1 only: cost of the source design under the target condition (106.73)	106.732
best_observed_feasible_cost	float64	USD/kWh	reference cost of the scenario, the lowest feasible cost over all primary evaluations (S1 97.104, S2 97.398, S3 100.044, S4 90.283)	90.2828, 97.1043, 97.3977, 100.044
normalized_gap_to_best	float64	fraction	final_best_feasible_cost / best_observed_feasible_cost - 1. Empty if no feasible design.	0 to 0.453444
terminal_status	string		success or failed	failed, success

configurations

Shared columns: scenario, strategy, optimizer, model_size, semantic_level.

Column	Type	Unit	Meaning	Values
n_feasible	int64		runs with feasible_found = 1	0 to 50
discovery_frac	float64	fraction	n_feasible / n_runs	0 to 1
median_first_feasible	float64		median first feasible evaluation over runs that found one	1 to 19
median_final_cost	float64	USD/kWh	median final best feasible cost over runs that found one	90.2828 to 117.789
protocol_failures	int64		runs that ended without a usable decision	0 to 8
cell_id	string		readable configuration key scenario|method|optimizer|size|level	248 distinct values
within2pct_pct	float64	%	near-best attainment, share of the 50 runs within 2% of the reference cost	0 to 100
within5pct_pct	float64	%	share of the 50 runs within 5%	0 to 100
iqr_final_cost_q1	float64	USD/kWh	first quartile of the final best feasible cost	90.2828 to 113.727
iqr_final_cost_q3	float64	USD/kWh	third quartile	90.7761 to 127.334
iqr_first_feasible_q1	float64		first quartile of the first feasible evaluation	1 to 16
iqr_first_feasible_q3	float64		third quartile	1 to 19.5

evaluations

Shared columns: run_id, scenario, x_chem, x_T, x_S, x_MPR, x_R.

Column	Type	Unit	Meaning	Values
k	int64		accepted BatPaC evaluation index, 1 to 20	1 to 20
proposal_source	string		what proposed the design: initial (warm start, source design or random start), ax, pymoo_ga, nomad_mads, random, llm_react	ax, initial, llm_react, nomad_mads, pymoo_ga, random
feasible	int64		1 if the design satisfies every target constraint	0, 1
violation_sum	float64	fraction	G of Eq. (D1), the sum of normalized positive constraint violations, as recorded	0 to 338.61
violation_max	float64	fraction	largest normalized constraint violation, derived after the runs	0 to 334.85
loss	float64	USD/kWh	penalized objective of Eq. (D1) passed to the optimizer	90.2828 to 3.4861e+07
cost	float64	USD/kWh	BatPaC pack cost per usable energy, the objective. Empty when the evaluation failed.	84.3669 to 354.593
total_pack_cost	float64	USD	BatPaC total pack cost	7537.78 to 33764.4
E	float64	kWh	pack energy	84, 130
P	float64	kW	pack power	239
V	float64	V	pack voltage	0.806411 to 5219
c	float64	Ah	cell capacity	4.02376 to 40302
L_bat	float64	mm	pack length	100 to 30758
W_bat	float64	mm	pack width	222.617 to 12768.4
fc_limit_um	float64	µm	S3 only: BatPaC's candidate-specific maximum cathode thickness for the 30-min recharge (Battery Design!G259)	-11126.1 to 107.467
fc_floor_um	float64	µm	S3 only: lower bound of that limit (Chem!C27)	15
applied_thickness_um	float64	µm	S3 only: cathode thickness applied by BatPaC (Battery Design!G269)	50 to 165
optimization_valid	int64		1 if the evaluation counts as valid. Invalid evaluations get the failure branch of Eq. (D1).	0, 1
failure_reason	string		missing_cost or simulator_exception for a failed evaluation	missing_cost, simulator_exception
wall_time_s	float64	s	wall time of the evaluation. The median is 6.39 s.	6.62352e-05 to 284.145
batpac_time_s	float64	s	time inside the BatPaC call (write, recalculate, read)	4.97916 to 139.521

vf_decisions

Shared columns: run_id, scenario, optimizer, model_size, semantic_level, seed.

Column	Type	Unit	Meaning	Values
partition_valid	int64		1 if the decision passed validation. 181 runs (119 at 4B, 62 at 9B) ended without a valid decision.	0, 1
mask	string		fixing mask, five letters in the order x^Chem, x^T, x^S, x^MPR, x^R. R = released to the optimizer, F = fixed at the source value	20 distinct values
fixed_x_chem	float64		1 if x_chem is fixed at its source value. Empty for invalid partitions.	0, 1
fixed_x_T	float64		1 if x_T is fixed at its source value. Empty for invalid partitions.	0, 1
fixed_x_S	float64		1 if x_S is fixed at its source value. Empty for invalid partitions.	0, 1
fixed_x_MPR	float64		1 if x_MPR is fixed at its source value. Empty for invalid partitions.	0, 1
fixed_x_R	float64		1 if x_R is fixed at its source value. Empty for invalid partitions.	0, 1
n_fixed	int64		fixed variables	0, 1, 2, 3, 4
n_free	int64		released variables	0, 1, 2, 3, 4, 5
required_released	float64		1 if every required variable of the scenario is released (S1 x_chem, S2 x_MPR and x_R, S3 x_T, S4 x_MPR)	0, 1
n_extra_released	float64		released variables outside the required set	0, 1, 2, 3, 4
n_required_released	float64		number of required variables released	0, 1, 2

ws_candidates

Shared columns: run_id, scenario, optimizer, model_size, semantic_level, seed, x_chem, x_T, x_S, x_MPR, x_R.

Column	Type	Unit	Meaning	Values
generation_id	string		identifier of the generation request	4,692 distinct values
candidate_idx	int64		candidate number 1 to 3	1, 2, 3
n_vars_changed_from_source	int64		variables whose value differs from the source design	0, 1, 2, 3, 4, 5
exact_source	int64		1 if the candidate equals the source design in all five variables	0, 1
in_target_region	int64		1 if the candidate moves in the scenario's target direction (S1 LFP, S2 x_R=3 and x_MPR>2, S3 x_T<155, S4 x_R=4 and x_MPR>2). A proxy, not feasibility.	0, 1
source_distance	float64	fraction	normalized distance from the source design	0 to 0.717209
initial_feasible	float64		1 if BatPaC evaluated the candidate as feasible	0, 1
initial_cost	float64	USD/kWh	BatPaC cost of the candidate	85.4489 to 241.48
in_empirical_topology_envelope	float64		S2 and S4 only: 1 if (x_R, x_MPR) lies in the observed feasible topology cells	0, 1

react_steps

Shared columns: run_id, scenario, model_size, semantic_level, seed.

Column	Type	Unit	Meaning	Values
k	int64		accepted BatPaC evaluation index, 1 to 20	1 to 20
reasoning_engineering	int64		1 if the text mentions an engineering mechanism, by the rules of Table 10 of the paper	0, 1
reasoning_constraint	int64		1 if the text targets a constraint	0, 1
reasoning_objective	int64		1 if the text targets the objective or cost	0, 1
reasoning_opt_history	int64		1 if the text refers to the optimization history with numeric feedback	0, 1
reasoning_source_ref	int64		1 if the text refers to the source design	0, 1
reasoning_exploration_uncertainty	int64		1 if the text expresses exploration or uncertainty	0, 1
reasoning_generic_unspecific	int64		1 if the text is generic and mentions none of the six other categories	0, 1
scenario_relevant_action	int64		1 if the step changes a variable relevant to the scenario	0, 1
edit_distance_norm	float64	fraction	normalized change from the accepted design before it (chemistry change and range-normalized variable changes, averaged)	0 to 0.779794
n_vars_edited	int64		variables changed from the accepted design before it	0, 1, 2, 3, 4, 5
coded_edit_size_source	int64		edit size relative to the source design as coded	0, 1, 2, 3, 4
feasible_after	int64		1 if the step's design is feasible	0, 1
pre_or_post_feasibility	string		phase of the step: initial/no predecessor, pre-feasible, transition-to-first-feasible, post-feasible	initial/no predecessor, post-feasible, pre-feasible, transition-to-first-feasible
label_count	int64		number of reasoning categories the text mentions	0 to 6
ambiguous_flag	int64		1 if the step has more than one label	0, 1
step	int64		agent step number, 1 to 60	1 to 60
reasoning_text	string		the reasoning text of the step, verbatim	21,963 distinct values
had_feasible_before	int64		1 if a feasible design was found before the step	0, 1

vf_rationale

Shared columns: scenario, optimizer, model_size, semantic_level, seed, run_id.

Column	Type	Unit	Meaning	Values
mask	string		fixing mask, five letters in the order x^Chem, x^T, x^S, x^MPR, x^R. R = released to the optimizer, F = fixed at the source value	20 distinct values
rationale_text	string		the rationale text returned by the model, verbatim	4,519 distinct values
reasoning_opt_history	int64		1 if the text refers to the optimization history with numeric feedback	0, 1
reasoning_engineering	int64		1 if the text mentions an engineering mechanism, by the rules of Table 10 of the paper	0, 1
reasoning_constraint	int64		1 if the text targets a constraint	0, 1
reasoning_objective	int64		1 if the text targets the objective or cost	0, 1
reasoning_source_ref	int64		1 if the text refers to the source design	0, 1
reasoning_exploration_uncertainty	int64		1 if the text expresses exploration or uncertainty	0, 1
reasoning_generic_unspecific	int64		1 if the text is generic and mentions none of the six other categories	0, 1
label_count	int64		number of reasoning categories the text mentions	0, 1, 2, 3, 4, 5
rationale_chars	int64	characters	length of the rationale text	0 to 524
text_available	int64		1 if a non-empty rationale text was recorded	0, 1

ws_rationale

Shared columns: scenario, optimizer, model_size, semantic_level, seed, run_id.

Column	Type	Unit	Meaning	Values
candidate_idx	int64		0 for the set-level allocation rationale, 1 to 3 for a candidate rationale	0, 1, 2, 3
rationale_text	string		the rationale text returned by the model, verbatim	18,972 distinct values
reasoning_opt_history	int64		1 if the text refers to the optimization history with numeric feedback	0, 1
reasoning_engineering	int64		1 if the text mentions an engineering mechanism, by the rules of Table 10 of the paper	0, 1
reasoning_constraint	int64		1 if the text targets a constraint	0, 1
reasoning_objective	int64		1 if the text targets the objective or cost	0, 1
reasoning_source_ref	int64		1 if the text refers to the source design	0, 1
reasoning_exploration_uncertainty	int64		1 if the text expresses exploration or uncertainty	0, 1
reasoning_generic_unspecific	int64		1 if the text is generic and mentions none of the six other categories	0, 1
label_count	int64		number of reasoning categories the text mentions	0 to 6
rationale_chars	int64	characters	length of the rationale text	0 to 5154
text_available	int64		1 if a non-empty rationale text was recorded	0, 1

react_calls

Shared columns: run_id, scenario, model_size, semantic_level, seed.

Column	Type	Unit	Meaning	Values
n_llm_calls	int64		generation calls of the run	20 to 61
n_rejected_proposals	int64		proposals returned as rejected (repeated or out of bounds)	0 to 58
n_invalid_json	int64		replies that could not be parsed	0, 1, 2
tokens_total	float64	tokens	tokens of the run	212789 to 1.16402e+06
n_accepted_evals	int64		accepted BatPaC evaluations	2 to 20
budget_filled	int64		1 if the run used all 20 evaluations	0, 1

react_call_events

Shared columns: run_id.

Column	Type	Unit	Meaning	Values
event_index	int64		position in the trace	0 to 61
event_type	string		llm_step, react_environment_step, react_terminal	llm_step, react_environment_step, react_terminal
accepted	bool		True if the proposed action was accepted	False, True
generation_attempts	int64		generation calls of the event	0, 1
explicit_rejection	int64		1 if the environment rejected the proposal	0, 1
invalid_json	int64		1 if the reply could not be parsed	0, 1

grid_evaluations

Shared columns: x_chem, x_T, x_S, x_MPR, x_R.

Column	Type	Unit	Meaning	Values
condition	string		source (the source condition, Table 6), S1 (the S1 cost condition), S3 or S4 (re-evaluation, Table 7)	S1, S3, S4, source
timestamp	string	ISO 8601	source and S1 rows only: time of the evaluation	2,748 distinct values
condition_name	string		condition name in the records: cost_84_239_base_long (source), cost_84_239_cobalt4x_long (S1), s_fastcharge (S3), s_energy130_relaxed_pack (S4)	cost_84_239_base_long, cost_84_239_cobalt4x_long, s_energy130_relaxed_pack, s_fastcharge
design_id	string		readable design key	1,374 distinct values
energy_req_kwh	float64	kWh	source and S1 rows only: required energy	84
power_req_kw	float64	kW	source and S1 rows only: required power	239
Lmax	float64	mm	source and S1 rows only: pack length limit	2200
Wmax	float64	mm	source and S1 rows only: pack width limit	1500
wall_time_s	float64	s	source and S1 rows only: wall time of the evaluation	4.7976 to 5.50593
batpac_time_s	float64	s	source and S1 rows only: time inside the BatPaC call	4.76294 to 5.46415
cost	float64	USD/kWh	BatPaC pack cost per usable energy, the objective. Empty when the evaluation failed.	87.2975 to 127.652
total_pack_cost	float64	USD	BatPaC total pack cost	7892.1 to 13101.7
E	float64	kWh	pack energy	84, 130
P	float64	kW	pack power	239
battery_capacity_ah	float64	Ah	source and S1 rows only: pack capacity	134.125 to 419.87
c	float64	Ah	cell capacity	33.5313 to 239.893
V	float64	V	pack voltage	135.477 to 626.28
mass_kg	float64	kg	source and S1 rows only: pack mass	386.05 to 673.889
volume_l	float64	L	source and S1 rows only: pack volume	225.608 to 447.975
L_bat	float64	mm	pack length	897.2 to 2532.4
W_bat	float64	mm	pack width	696.858 to 2382.9
fc_limit_um	float64	µm	S3 only: BatPaC's candidate-specific maximum cathode thickness for the 30-min recharge (Battery Design!G259)	85.455 to 102.476
fc_floor_um	float64	µm	S3 only: lower bound of that limit (Chem!C27)	15
feasible	float64		S3 and S4 rows only: 1 if feasible under the scenario	0, 1
penalized_loss	float64	USD/kWh	S3 and S4 rows only: Eq. (D1)	90.2828 to 1.166e+06

l2_comprehension_responses

Shared columns: scenario.

Column	Type	Unit	Meaning	Values
repetition	int64		1 to 3	1, 2, 3
source_file	string		file name in the archive l2_comprehension	12 distinct values
response_chars	int64	characters	length of the response	10269 to 13609
response_text	string		the response, verbatim	12 distinct values

Archives

Archive	Content
archives/result_tables.tar.zst	the result tables in _pack_staging/04_results_tables/ and the scenario specification, which the code of the paper reads
archives/preliminary_grid.tar.zst	the preliminary design grid of Appendix A and its evaluation under S3 and S4
archives/runs_qwen35_4b.tar.zst	the raw records of the 3,600 runs with Qwen3.5-4B
archives/runs_qwen35_9b.tar.zst	the raw records of the 3,600 runs with Qwen3.5-9B
archives/runs_qwen35_27b.tar.zst	the raw records of the 3,600 runs with Qwen3.5-27B
archives/runs_numerical.tar.zst	the raw records of the 1,600 runs of the cold start and the source-warm start

A run folder holds the prompts, the raw replies, the request trace (trace.jsonl), every BatPaC evaluation (real_batpac_records.jsonl) and the run summaries. Machine paths in the records appear as placeholders such as <server-home>.

License and citation

The data are released under CC BY 4.0. The code is under the MIT license.

@unpublished{song_wafer,
  author = {Song, Chan-Eui and Michalek, Jeremy J. and McComb, Christopher},
  title  = {{LLM}-Assisted Optimization for Simulation-Based Redesign},
  note   = {Submitted to the ASME Journal of Mechanical Design (2026)},
  year   = {2026}
}

BatPaC

BatPaC is developed by Argonne National Laboratory and distributed under a BSD license. The workbook is not included. Obtain BatPaC 5.0 (2022-02-09) from https://www.anl.gov/cse/batpac-model-software and cite its manual:

K. W. Knehr, J. J. Kubal, P. A. Nelson, and S. Ahmed, "Battery Performance and Cost Modeling for Electric-Drive Vehicles: A Manual for BatPaC v5.0," Argonne National Laboratory, ANL/CSE-22/1, July 2022. doi:10.2172/1877590.