PaperlayerView exact revision on arXiv

Source-linked AI summary

EvoSim: Learning to Model, Modeling to Learn

Yun-Wei Song, Jinkai Tao, Jun-Dong Zhang, Rui Zhang, Yi-Min Wu, Qiang Zhang

arXiv:2610.11344v1cs.AIcs.CE

TL;DR

Existing AI systems remain limited in autonomously choosing the structures of physics-based models, a task that requires selecting mechanisms, states, equations, couplings, and parameters. EvoSim revises candidate models using experimental discrepancies and learns research capabilities from investigation records. Across two industrial battery tasks, it reports predictive accuracy and fewer model-construction or physical-mechanism failures with capability learning.

  • Problem

    Existing AI systems remain limited in autonomously choosing physical-model structures, although construction requires selecting processes, states, equations, couplings, and parameters while preserving physical constraints.

    Source: §1 · p. 1

  • Method

    EvoSim explores model changes in parallel using experimental discrepancies, validates candidates against measurements, and extracts evidence-checked knowledge and procedures from investigations for later use.

    Source: §1 · p. 2, §5 · p. 10

  • Results

    Across two industrial battery tasks, EvoSim reported 1.79% points in state of charge MAE for plating onset, 7.62 mV RMSE for dynamic voltage, and a 36.4% reduction in model-construction or physical-mechanism failures.

    Source: §1 · pp. 2–3, §5 · p. 10

  • Takeaways & Limitations

    Successful and failed investigations can both contribute reusable research expertise alongside the development of physics-based models.

    Source: §5 · pp. 9–10

Abstract

from arXiv · show

Physics-based models connect scientific explanation with quantitative prediction. Constructing them requires selecting physical processes, defining states and governing equations, specifying couplings, and identifying parameters from experiments. Existing AI systems remain limited in making these model structure decisions autonomously. We introduce EvoSim, a self-evolving AI scientist for physical modeling. It uses experimental discrepancies to drive mechanism and equation revisions and held-out experimental data to test physical plausibility. Exploration traces make updates to knowledge, skills, and multi-agent orchestration. This co-evolution improves physics-based models and EvoSim's ability to select mechanisms, diagnose failures, and coordinate research. We evaluate EvoSim on two industrial battery modeling tasks. It predicts lithium-metal-plating onset from 25 to 45 degrees Celsius and 2 C to 6 C with a mean absolute error of 1.79% in state of charge. Dynamic voltage prediction under vehicle driving conditions achieves a root mean square error of 7.62 mV, surpassing the reported accuracy of models developed by human experts. Self-evolution reduces model and physics errors by approximately 36% relative to baseline, demonstrating improved scientific modeling capability. EvoSim turns experimental observations into validated models and cumulative research expertise.

1 INTRODUCTION

EvoSim addresses the human-defined model-space constraint by revising physical formulations from experimental discrepancies while learning reusable research capabilities. Its industrial battery evaluations report predictive accuracy and fewer model-construction or physical-mechanism failures with capability learning.

1 INTRODUCTION

  • Physics-based modeling requires selecting processes, states, equations, couplings, and parameters while maintaining physical constraints, and unknown mechanisms make those choices harder.Existing AI systems usually operate within human-specified model spaces and cannot autonomously construct new explanatory and predictive models.Source: §1 · pp. 1–2
  • Fixed workflows and preloaded knowledge are limited because relevant knowledge and action sequences must evolve as new evidence changes the modeling problem.EvoSim reviews investigation records to extract domain knowledge, modeling skills, tool procedures, and research strategies for later use.Source: §1 · p. 2
  • 1.79% points in state of charge MAE for lithium-plating onset and 7.62 mV RMSE for dynamic voltage were reported across two industrial battery tasks.Across 45 matched comparisons, capability learning reduced model-construction or physical-mechanism failures from 73.3% to 46.7%.Source: §1 · p. 2
  • EvoSim couples experimental model development with research-capability learning, so model investigations identify learning needs and retained methods support further exploration.A rejected model can still yield a useful research method.Source: §1 · p. 2
  • It investigates competing explanations for prediction discrepancies in parallel by changing mechanisms, equations, and couplings, then checks candidates against measurements.Candidate results guide further exploration and progressive refinement through experimental feedback.Source: §1 · p. 2
  • 36.4% fewer model-construction or physical-mechanism failures and runtime reduced to 57.4% of the raw condition were reported across 45 matched comparisons.A voltage-modeling task also reached 22.015 mV RMSE within three submissions, reducing error by 75.4% versus chronological summaries; that baseline did not reach comparable accuracy after twelve submissions.Source: §1 · p. 3

2 RELATED WORK

EvoSim differs from systems that execute within a predefined model space by making physical formulations revisable. It also couples model development to evidence-checked updates of research capabilities.

2.1 AUTONOMOUS SCIENTIFIC RESEARCH

  • EvoSim makes the formulation itself revisable, testing competing physical explanations through changes to equations, states, and couplings while revising investigation plans.Its autonomy concerns both what model to build and what investigation to pursue, rather than simulation execution alone.Source: §2.1 · p. 3

2.2 SCIENTIFIC EVOLUTION AND SELF-IMPROVEMENT

  • Scientific Evolution distinguishes observed failures, proposed corrections, and procedures supported by application, allowing rejected models to improve research methods without endorsing their explanations.Updated methods guide later hypotheses, model construction, and investigation plans, whose outcomes inform further updates to both model and scientist.Source: §2.2 · p. 3

3 METHOD

EvoSim coordinates model and capability updates through a research loop. It tests planned model changes against measurements and uses investigation evidence to revise the scientist’s reusable knowledge and methods.

3.1 ARCHITECTURE

  • EvoSim couples model development with research-capability updates, using investigation records to coordinate work and retained scientist methods to guide later research.The research state records hypotheses, branches, completed investigations, evidence, and open questions; the scientist state stores findings, failure experience, and methods for later use.Source: §3.1 · pp. 3–5, Figure 2 · p. 4

3.2 MODEL EVOLUTION

  • Model evolution proposes explanations for residuals, organizes focused or parallel investigations, and validates candidate implementation, numerical solutions, and agreement with measurements.Reviewed evidence informs model selection and later rounds, including when a candidate is rejected; exploration traces connect proposed changes to implementation and results.Source: §3.2 · p. 5

3.3 SCIENTIFIC EVOLUTION

  • Scientific Evolution updates domain knowledge, modeling and validation procedures, and research strategies from reviewed evidence and exploration traces.Failures guide diagnosis, while proposed corrections become reusable procedures only when application supports their expected effect; retained strategies are reassessed in later investigations.Source: §3.3 · pp. 5–6

4 EXPERIMENTS

EvoSim was evaluated on industrial battery predictions and controlled studies of whether scientific evolution improves later modeling. The experiments assess predictive accuracy, reliability, efficiency, and the agreement of scientific conclusions with evaluation results.

4.1 EXPERIMENTAL SETUP

  • The evaluation studies two industrial battery tasks and uses implementation, solution, and measurement checks on each model.The implementation check verifies active equations and states; the solution check verifies completion and finite outputs; the measurement check compares predictions with specified observations.Source: §4.1 · p. 6

4.2 INDUSTRIAL BATTERY MODELING

  • 7.62 mV for 30°C DST and 8.97 mV for 20°C US06 are the reported voltage cross-validation errors.Figure 3c shows complete measured and predicted traces for these two cross-validation results.Source: Figure 3 · p. 7
  • 1.79 SOC percentage points MAE and 2.11 SOC percentage points RMSE were achieved on six held-out lithium-plating onset conditions.The six validation conditions were excluded from parameter identification.Source: §4.2 · p. 8
  • Voltage evaluation leaves out one temperature at a time, using other temperatures for parameter identification and comparing complete predicted and measured traces.Figure 3c shows traces at two held-out temperatures.Source: §4.2 · p. 8

4.3 EFFECT OF SCIENTIFIC EVOLUTION

  • On the 30°C DST task, Scientific Evolution reached its best accepted model on attempt 2, with experimental-voltage RMSE 75.4% lower than Baseline; Naive Memory stayed close to Baseline.The comparison held task, starting model, data, solver, tools, checks, and maximum attempts constant.Source: §4.3 · p. 8
  • Saved procedures reduced required attempts from 8 to 2 in capacity matching and model changes from 10 to 1 in P2D–thermal modeling.In the thermal-mechanism task, the comparison procedure accompanied a temperature RMSE decrease from 9.30 to 3.74°C.Source: §4.3 · p. 8
  • 33 to 21 model and physical-mechanism failure runs across 45 matched pairs; Scientific Evolution also reduced active execution time by 42.6%.Among pairs with different outcomes, 15 favored Scientific Evolution and 3 favored Baseline; the two-sided exact McNemar test gave p = 0.0075, with p = 0.0151 after adjustment across six failure modes.Source: Figure 4 · p. 9, §4.3 · p. 9
  • All nine scientific-conclusion decisions agreed with evaluation results, including rejecting one proposed general mechanism after every tested model failed the evaluation criterion.Four conclusions applied to tested conditions and four applied only to narrower ranges.Source: §4.3 · p. 9

5 CONCLUSION

EvoSim co-evolves physics-based models with reusable modeling capabilities, using evidence from investigations to update knowledge and procedures for subsequent work. The reported battery tasks show predictive results, while the evaluation setup fixes task inputs and acceptance rules.

5 CONCLUSION

  • EvoSim co-evolves physics-based models and reusable modeling capabilities, extracting knowledge and procedures from successful and failed attempts for subsequent investigations.The evolving process checks supporting evidence before applying retained knowledge, procedures, and research strategies.Source: §5 · pp. 9–10
  • The battery plating-onset model achieves 1.79% MAE in state of charge, the dynamic-voltage model reaches 7.62 mV RMSE, and failures decrease by 36.4% across 45 matched comparisons.The paper presents these industrial-task results as evidence of improvements in model accuracy, modeling efficiency, and reliability.Source: §5 · p. 10

A MODELING SETUP AND EVALUATION

  • Each experiment fixes the cell, protocol, observations, starting equations, permitted modifications, solver, output quantities, comparison metric, and stopping rule before the agent begins.Candidate modifications can be explored serially, in parallel, or both, and research records link modifications to evidence and decisions.Source: A MODELING SETUP AND EVALUATION · p. 13

A.1 CONTROLLED MODELING PROBLEM

  • A run specifies the task, executable starting model, observation data, simulator, permitted changes, work budget, and fixed acceptance rules; some tasks also reserve separate data for final comparison.These inputs determine what the agent can read, change, and must produce.Source: §A.1 · p. 13
  • A research record saves a candidate modification or set of modifications with its rationale, result, and decision to keep or reject it.The fixed task inputs include the task description, starting-model source, data, required outputs, checks, and budget.Source: §A.1 · p. 13

A.2 STARTING PHYSICS-BASED MODEL AND FIXED INPUTS

  • The physical experiment supplies the cell, measurement protocol, instruments, and records; the modeling setup specifies the executable representation, allowed edits, outputs, and stopping rule.Measurements remain the reference for constructing and evaluating the model.Source: §A.2 · p. 13

A.3 MODELING AGENT AND SOLVER EVALUATION

EvoSim explores executable battery-model candidates against fixed experimental tasks, records each investigation, and uses validated procedures to guide later modeling. Its evaluation separates implementation, numerical solution, measurement agreement, and physical checks.

A.4 ISOLATION AND SAVED MODEL RECORDS

  • Candidates are isolated and linked to their model changes, parameters, solver evidence, comparisons, diagnosis, and decision; rejected candidates remain in the record without replacing the selected model.Earlier records are available only when the experiment allows them.Source: §A.4 · p. 14

A.5 EVALUATION CHECKS, BUDGET, AND TERMINATION

  • Scientific Validation checks implementation, successful solution, and agreement with measurements; a candidate replaces the current model only after all required checks pass.Implementation checks that required physics and states are active; solution checks completion and finite, valid outputs, while measurement checks use the fixed metric.Source: §A.5 · pp. 14–15

B.1 RESEARCH PROCEDURE

  • Each exploration step records a candidate’s parent, changes, expected effect, observations, outputs, metric, acceptance rule, results, diagnosis, and decision.Candidates may run serially or in parallel, and later steps can continue from a saved model or selected branch.Source: §B.1 · p. 15

B.2 REUSABLE PROCEDURES AND LATER USE

  • A saved procedure carries its result, action, expected response, checks, and supported conditions; later evidence can extend, restrict, or replace it.If new evidence does not support revision, the existing procedure remains unchanged.Source: §A.3 · p. 14, §B.2 · p. 15

C.1 ROLE OF MEASUREMENTS IN PHYSICS-BASED MODEL DEVELOPMENT

  • Measurements constrain equations, identify parameters, and quantify predictions, while solution and physical checks test equation execution and constraints.The plating case uses 14 training and six fixed validation conditions; dynamic voltage uses leave-one-temperature folds on complete driving traces.Source: §C.1 · pp. 15–16

C.2 SHARED MODEL DEVELOPMENT SEQUENCE

  • The battery cases follow distinct model-development paths: plating adds a missing reaction to a DFN/P2D foundation, while voltage extends a four-state baseline with electrolyte transport and interfacial charging.The plating path connects deposited lithium to the measured onset definition.Source: §C.2 · p. 16, §C.3 · p. 16

C.3.1 SCIENTIFIC EXPLORATION AND RETAINED MODEL

The lithium-plating model combines a DFN/P2D cell description with plating/stripping and a cumulative irreversible-lithium estimate. EvoSim tests a rate-dependent inactivation yield using training conditions, then evaluates onset against held-out conditions.

Experimental matrix and target

  • The 20-condition matrix spans two electrode designs, three temperatures, and 2–6 C charging; 14 conditions train the model and six remain fixed for validation.Onset is the first SOC at which accumulated irreversible lithium reaches 0.05% of graphite capacity.Source: §C.3.1 · pp. 17–18

Model exploration

  • A plating/stripping reaction predicts earlier onset at faster rates and for the thicker A2 electrode, while deposition is accumulated separately because instantaneous current does not match the irreversible-lithium onset label.The deposition-only state retains deposited charge even when reversible plated lithium later decreases during stripping.Source: §C.3.1 · p. 18

Model comparison

  • EvoSim compares a constant 0.30 irreversible fraction with a rate-dependent yield, keeping geometry, transport, diffusion, and intercalation fixed.The rate-dependent parameters are selected using only the 14 training conditions, with the six validation conditions excluded until the model is fixed.Source: §C.3.1 · pp. 18, 19, 21

Retained model

  • The retained model couples electrolyte salt and charge conservation, spherical-particle lithium diffusion, graphite intercalation, and a parallel plating/stripping reaction.Transport functions and intercalation kinetics are implemented and verified before plating parameters are identified, and remain fixed during the kinetics comparison.Source: §C.3.1 · pp. 19–20

Irreversible-lithium representation

  • The model distinguishes gross deposited lithium from the fraction becoming irreversible within the measurement window, representing their connection through an effective inactivation yield.The rate dependence is a bounded low-order closure over the measured 2–6 C range; its exponent also absorbs unresolved effects of morphology, contact loss, and stripping accessibility.Source: §C.3.1 · pp. 20–21

Rate dependence

  • The rate-dependent yield represents competition between charging time and lithium inactivation time; n = 1 corresponds to inverse charging-time scaling alone, while the identified n = 1.75 also absorbs unresolved rate effects.Higher charging rates can increase deposition while reducing the fraction of each deposited coulomb that becomes inactive within the observation window.Source: §C.3.1 · p. 21

C.3.2 CELL-SPECIFIC ONSET IDENTIFICATION

On the six held-out plating-onset conditions, EvoSim’s rate-dependent model has lower reported errors than the compared source-study model, whose parameters were developed using all 20 conditions. The section also introduces a separate dynamic-voltage evaluation on complete driving traces.

C.3.2 CELL-SPECIFIC ONSET IDENTIFICATION

  • The rate-dependent representation has validation MAE 1.79 percentage points and RMSE 2.11 percentage points, compared with 3.79 and 5.03 percentage points for the constant-fraction representation.The rate-dependent parameters were identified only on the 14 training conditions.Source: §C.3.2 · pp. 21–22, Table 7 · p. 22
  • 1.79 percentage points MAE and 2.11 percentage points RMSE are EvoSim’s six-condition validation errors, versus source-study values of 2.82 and 4.06 percentage points, respectively.EvoSim excluded those six conditions from parameter identification; the source-study model used the full 20-condition matrix during development.Source: §C.3.2 · p. 22, Table 8 · p. 23

C.4 DYNAMIC VOLTAGE PREDICTION

  • The dynamic-voltage task evaluates complete DST and US06 traces across temperatures using leave-one-temperature cross-validation on 16 traces.The supplied passage describes the evaluation setup, not its numerical result.Source: §C.4 · p. 24

C.4.1 SCIENTIFIC TASK AND EVALUATION PROTOCOL

The extended voltage model adds electrolyte transport and interfacial charging and reports two held-out run results. The evaluation distinguishes complete-data fitting from temperature-held-out testing.

C.4.1 SCIENTIFIC TASK AND EVALUATION PROTOCOL

  • The 16 public DST and US06 traces cover a 1.1 Ah LiFePO4 cell measured across eight temperatures, with complete voltage sequences under dynamic loads.The traces include rapid load changes, current reversals, rest periods, and voltage recovery; the task tests state tracking through complete load sequences.Source: §C.4.1 · p. 24
  • Eight leave-one-temperature folds test transfer to an unseen temperature within the represented DST and US06 protocols, not to an unseen drive protocol.Each fold identifies parameters using 14 traces at the other seven temperatures, then predicts both complete traces at the held-out temperature without refitting.Source: §C.4.1 · p. 24

C.4.2 SOLVED ELECTROCHEMICAL MODEL

  • The four-state baseline uses exact one-second zero-order-hold state updates and temperature-dependent transport, kinetics, and internal resistance.Its four states are the average and surface filling fractions of the negative and positive electrodes.Source: §C.4.2 · p. 25

C.4.3 PHYSICAL EXTENSIONS BEYOND THE FOUR-STATE BASELINE

  • The extensions add electrolyte transport and interfacial charging to the solved equations, connecting these processes with solid diffusion and ohmic loss.Current-step, early transient, sustained-load, relaxation, and multi-temperature data constrain different parameter groups.Source: §C.4.3 · pp. 26–27

C.4.4 MODEL DEVELOPMENT SEQUENCE

  • 7.62 mV on the 30°C DST trace and 8.97 mV on the 20°C US06 trace are held-out results from the extended model.The four-state baseline gives 32.12 mV equal-condition mean RMSE across 16 public DST and US06 traces; the reported held-out values are individual runs.Source: §C.4.2 · p. 26, §C.4.4 · p. 27

C.4.5 COMPLETE-TRACE DEVELOPMENT CALCULATIONS AND CROSS-VALIDATION RESULTS

The appendix reports complete-data fits separately from the individual temperature-held-out voltage results. The section also frames Scientific Evolution evaluation around controlled comparison, recursive updating, and later-task use.

C.4.5 COMPLETE-TRACE DEVELOPMENT CALCULATIONS AND CROSS-VALIDATION RESULTS

  • 4.86 mV is the equal-condition mean RMSE for fitting all 16 traces, not a held-out result.The appendix panels show complete runs using all traces for parameter identification and evaluation; Figure 3c separately shows two traces at held-out temperatures.Source: §C.4.5 · p. 28

C.4.6 REFERENCE MODELS

  • The 25°C DST reference comparison reports 6.08 mV for the physics-based run and 32.10 mV for published CPG-SPMT, using each study’s calibration procedure.The 6.08 mV physics-based result is not held out.Source: §C.4.6 · p. 28

C.4.7 COMPARISON WITH PUBLISHED VOLTAGE MODELS

  • Table 10 organizes dynamic-voltage RMSE values by DST or US06 condition, separating a same-condition reference block from designated cross-validation results and literature values.Its first block shares the cell and measurement condition while retaining each study’s original calibration procedure.Source: §C.4.7 · p. 28, Table 10 · p. 29

D SCIENTIFIC EVOLUTION COMPARISONS

  • Scientific Evolution is evaluated through a controlled next-decision comparison, a four-step recursive record, related-task use, and 45 paired experiments.The four-step record tracks how a model result changes a procedure, which then guides a later investigation and another update.Source: D SCIENTIFIC EVOLUTION COMPARISONS · p. 31

D.1 CONTROLLED TEST OF THE SCIENTIFIC EVOLUTION MODULE

The controlled comparison tests whether a checked procedure changes subsequent model decisions relative to a chronological summary or no prior record. Further examples trace procedure reuse in related tasks and successive evidence-driven updates.

D.1 CONTROLLED TEST OF THE SCIENTIFIC EVOLUTION MODULE

  • The matched comparison gives each condition the same task, starting model, data, solver, tools, and checks; the procedure records an action, its applicable conditions, and checks.Baseline gets no earlier record, while Naive Memory receives a chronological summary and Scientific Evolution a checked procedure from the same investigation.Source: §D.1 · pp. 31–32
  • The checked procedure changes the sequence of model changes and reaches a lower-error valid model, while the chronological summary remains near the baseline result.The procedure condition reaches its lower-error valid model at B = 2; Baseline and Naive Memory use further submissions without improving their best valid results.Source: §D.1 · pp. 32–33, Table 13 · p. 32, Table 14 · p. 33

D.2 USE AFTER SCIENTIFIC EVOLUTION

  • Three related-task records test procedure reuse with each target task’s own model, data, and acceptance check; the cases span revision effort, multiphysics coupling, and mechanism selection.Together with the voltage case, the four outcomes cover implementation accuracy, revision effort, coupling, and mechanism selection.Source: §D.2 · pp. 34–35, Table 18 · p. 35

D.3 HOW SCIENTIFIC EVOLUTION PRODUCES A PROCEDURE

  • Across four successive investigations, a failed model informs a correction, results across 24 conditions expand a procedure’s supported conditions, and later evidence updates it again.The recorded result changes the procedure used in each subsequent step.Source: §D.3 · p. 35, Table 19 · p. 36

D.4 EVALUATION ON LATER TASKS AFTER SCIENTIFIC EVOLUTION

Scientific Evolution reduces recurring model and physics failures in matched later-task runs, while the appendix also specifies the evaluation scope, task collection, and cross-task conclusions.

D.4 EVALUATION ON LATER TASKS AFTER SCIENTIFIC EVOLUTION

  • 36.4% relative reduction in model / physics failures (33/45 to 21/45) remained significant after Holm correction across six failure modes.False completion, data/protocol failure, tool-operation failure, and looping/stagnation also remained significant; numerical-solver failure did not show a significant difference.Source: §D.4 · p. 37, Table 24 · p. 38

D.4.1 EXECUTION EFFICIENCY

  • Active runtime decreased by 42.6% in the paired later runs; totals exclude the earlier investigations that produced and checked the procedures.The paired analysis uses 45 pairs across 11 tasks, with the same starting model, data, solver, tools, checks, rules, and budget within each pair.Source: §D.4 · p. 36, §D.4.1 · p. 38

E TASK COLLECTION

  • The task collection distinguishes fixed-model evaluation from testing whether a learned procedure changes decisions on a later task.The evaluation applies a fixed model to another condition or dataset; Section D.2 addresses procedure transfer to later tasks.Source: E TASK COLLECTION · p. 38

E.1 TASK INVENTORY

  • The inventory spans parameter identification, forward prediction, mechanism and multiphysics studies, and model diagnosis.These categories distinguish choosing model values, applying a fixed model, testing physical processes and couplings, and recording why a change failed.Source: §E.1 · p. 39

E.2 RUN ACCOUNTING

  • The collection contains 179 runs in 32 task families; cross-task analysis uses 65 records from 26 families, including nine with complete comparison chains.A complete chain has a candidate fixed before evaluation, a prespecified target, a numerical comparison, and a final decision.Source: §E.2 · p. 39

E.3 CROSS-TASK EVALUATION OF SCIENTIFIC DECISIONS

  • Nine complete records follow fixed models through later comparisons and final decisions; four conclusions hold for evaluated conditions, four are narrowed, and one proposed general mechanism is rejected.The tables report the comparison and conclusion for each record.Source: §E.3 · p. 39

E.3.1 TASK-FAMILY SCREENING

The screening identifies nine task families with complete evaluation records, and the included chains use comparison outcomes to set the scope of scientific conclusions.

E.3.1 TASK-FAMILY SCREENING

  • Nine of 26 task families contain a complete evaluation record with a fixed candidate, comparison target, numerical result, and final decision.The analysis retains one complete record per family and checks whether its numerical result agrees with the recorded decision.Source: §E.3.1 · p. 39
  • The screening table separates complete evaluations from model-development and partial records.The nine complete records support the ensuing summary.Source: §E.3.1 · p. 40, Table 29 · p. 40

E.3.2 INCLUDED EVALUATION CHAINS

  • None of the tested negative solid-electrolyte interphase candidates passed the comparison across all stages and cells, so the proposed general mechanism was rejected.Decision labels indicate whether a conclusion is tested only, limited to supported conditions, or unsupported.Source: §E.3.2 · p. 41, Table 30 · p. 41
  • EvoSim defines a target, records and tests candidate modifications under relevant conditions, then uses the result to select a model and the next candidate.A universal mechanism must pass the stated comparison at every stage and in every cell covered by the conclusion.Source: §E.3.2 · p. 41

F REPRODUCIBILITY

The reproducibility account fixes model, solver, and evaluation details and records inputs and outputs; reusable procedures turn investigated results into checked modeling actions for later use.

F REPRODUCIBILITY

  • Controlled comparisons fix the language-model setup and research inputs; numerical comparisons also hold solver and model settings constant, and saved records identify each candidate and its result.The records include task and input hashes, repository revision, submitted model, numerical results, metric calculation, and review decision.Source: F REPRODUCIBILITY · pp. 41–42

G ADDITIONAL REUSABLE MODELING PROCEDURES

  • RMSE against the reference calculation decreases from 88.518 to 4.1136 mV after applying the discrete zero-order-hold state update once per sample.The same update solves 24 temperature and drive-cycle conditions with a mean reference-calculation RMSE of 4.4625 mV; later use against measured voltage gives 22.015 mV RMSE.Source: G ADDITIONAL REUSABLE MODELING… · p. 42
  • The reusable state update applies to one-second sampling, piecewise-constant current and temperature inputs, the stated CPG-SPMT equations, and voltage equation.The implementation converts temperature to Kelvin once and assigns four average and surface filling fractions to dependent variables.Source: G ADDITIONAL REUSABLE MODELING… · p. 42
  • To distinguish a missing process from a parameter error, first verify the proposed term enters the active equation and test an input with a clear predicted response direction.Change parameters only after the equation produces that response; the later decision is whether to add a process, repair a coupling, or continue parameter identification.Source: G ADDITIONAL REUSABLE MODELING… · p. 43
  • A solved model does not establish that an added reaction or coupling is active; trace its term into the relevant equation and confirm the target output changes as predicted.Retain the mechanism only when the term is active and its response passes the measurement check.Source: G ADDITIONAL REUSABLE MODELING… · p. 43
  • Separate heat-generation and heat-transfer candidates before combining mechanisms; the later P2D–thermal investigation reduced sequential model changes from 10 to 1.The procedure compares particle-size effects, polarization heat, boundary heat flux, and convection under the same input and temperature metric.Source: G ADDITIONAL REUSABLE MODELING… · p. 43
Loading 2610.11344v1…