Organizations: Beijing Key Laboratory of Complex Solid-State Batteries, Department of Chemical Engineering, Tsinghua University, Beijing, China · State Key Laboratory of Chemical Engineering and Low-Carbon Technology, Tsinghua University, Beijing, China · Beijing Tsingyu Technology Co., Ltd., Beijing, China · School of Computing and Data Science, The University of Hong Kong, Hong Kong, China · Beijing Huairou Laboratory, Beijing, China · Tanwei College, Tsinghua University, Beijing, China · AI Solid-State Battery Innovation Center, Yibin, China · Shanxi Research Institute for Clean Energy, Tsinghua University, Taiyuan, China
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.
Figures & tables
Figure 1: EvoSim co-evolves a physics-based model and the AI scientist that develops it.
Figure 2: EvoSim’s architecture and coupled evolution. (a) The research loop connects hypothesis formation, research planning, model construction, validation, and Scientific Evolution. (b) Model evolution develops and evaluates alternative hypotheses through computational modeling and scientific validation. Scientific evolution uses feedback from past investigations to update knowledge, skills, tools, and research strategies.
Figure 3: Lithium-plating onset and dynamic-voltage predictions. (a) Onset SOC and accumulated irreversible lithium for two graphite-electrode designs across temperatures and charging rates. (b) Measured and predicted onset SOC for 20 conditions (left), and voltage RMSE compared with reference values (right). The voltage panel reports cross-validation errors for 30 ∘ C DST (7.62 mV) and 20 ∘ C US06 (8.97 mV). (c) Complete measured and predicted voltage traces for the same two cross-validation results.
Figure 4: Effect of Scientific Evolution. (a) Baseline investigations and investigations using saved procedures on three tasks. (b) Failure counts and execution measures across 45 matched pairs from 11 modeling tasks. Each measure is normalized to its corresponding Baseline value, shown as 100%. The figure label Improved denotes EvoSim with Scientific Evolution.
Condition
Best valid RMSE (mV)
Progress during search
Baseline
89.478
Unchanged through 12 attempts
Naive Memory
89.383
Unchanged through 12 attempts
Scientific Evolution
22.015
Attained on attempt 2
Table 1: Ablation on the four-state voltage task. Performance and progress during the model search.
Appendix figures & tables31 assets
Supplementary material from the paper’s appendix.
Appendix
Part
Fixed before a run
Operation available to the agent
Evidence saved after the run
Fixed task inputs
Scientific target, starting model, input conditions, required outputs, observations, and run rules
Read the fixed files and propose permitted model modifications
Input-file list, file hashes, and task directory
Starting model
Physics-based equations, geometry, parameters, initial conditions, input time range, and relation to the observations
Copy the current model and edit permitted model or solver elements
Submitted Java source or physics-based model file, parameter values, and model-state checks
Multiphysics solver
Local or remote solver and the required output format
Build each candidate, solve it, and request the specified fields and states
Solver status, logs, output files, and resource use
Research record
Current model, permitted earlier records, saved procedures, and acceptance rule
Interpret feedback and select the next candidate under the same rules
Hypothesis, change, diagnosis, decision, and link to the parent model
Checks and metric
Implementation checks, solution checks, measurement metric, and any specified comparison target
Receive the observations and feedback released by the task
Rebuild, solve, export, compare, and record the decision
Budget and isolation
Work slots, time and job limits, fixed inputs, and permitted directories
Spend a slot and write only inside the candidate directory
Input-file list, access log, budget record, and saved outputs
Appendix
Table 2: Fixed inputs, permitted actions, and saved evidence in the reported experiments.
Case
Observed result
Physical hypothesis and model change
Evidence and decision
Plating
A basic DFN/P2D model requires cell-specific transport and reaction functions before onset can be calculated.
Specify geometry, electrolyte transport, particle diffusion, graphite equilibrium potential, and intercalation kinetics for A1 and A2.
Unit, reference-condition, positivity, state-limit, and conservation checks establish an executable Li ∣ graphite model.
Plating
Transport and intercalation cannot generate metallic lithium.
Add a competing plating/stripping reaction driven by ηLi=ϕs−ϕe−ULi .
Retain only if negative plating overpotential activates deposition during fast charging and the added reaction preserves lithium conservation.
Plating
Instantaneous reaction current and net plated lithium can decrease during stripping.
Integrate max[−asiLi,0] through the graphite thickness and time, then convert the accumulated deposition to irreversible lithium.
Retain the monotonic accumulated state because it matches the measured onset definition and remains well defined during later stripping.
Plating
A1 and A2 have different areal capacities and onset boundaries.
Normalize the onset threshold by areal capacity and represent irreversible lithium as a fixed fraction of accumulated deposition.
The capacity-normalized thresholds support comparison across the measured A1 and A2 conditions.
Plating
The constant-fraction representation retains systematic onset residuals across charge rates.
Compare it with a rate-dependent mapping from deposited lithium to irreversible lithium.
On the same six validation conditions, the rate-dependent representation lowers MAE from 3.79 to 1.79 percentage points and RMSE from 5.03 to 2.11 percentage points.
Voltage
Four electrochemical states reproduce the main voltage level but not the response after a current change.
Solve average and surface electrode states with temperature-dependent transport, kinetics, and resistance.
The starting diagnostic gives 32.74 mV.
Appendix
Table 3: Physical observations, tested hypotheses, and decisions in the two model-development paths.
Electrode
Areal capacity
Thickness
Temperature
Charge rates
(mAh cm -2 )
( μ m)
( ∘ C)
A1
2.1
47
25, 35, 45
4 C, 5 C, 6 C
A2
3.1
70
25
2 C, 3 C, 4 C, 5 C, 6 C
A2
3.1
70
35, 45
4 C, 5 C, 6 C
Appendix
Table 4: Experimental matrix used for lithium-plating onset modeling.
Observation
Tested model change
Decision
The common starting point supplies only the basic DFN/P2D equations.
Set electrode thickness, phase fractions, particle radius, initial concentrations, and the conversion from C-rate to applied current.
Retained after geometry, capacity, and initial-state checks for A1 and A2.
Electrolyte polarization changes strongly with concentration and temperature.
Implement De(ce,T) , κe(ce,T) , and t+0(ce,T) together with porous-medium corrections.
Verify finite positive transport properties and 0<t+0<1 over 25–45 ∘ C and the simulated concentration range.
Graphite transport and reaction rates change with filling fraction and temperature.
Implement Ds(x,T) , Ugr(x) , and i0,int(ce,cs,T) and evaluate them at the reference conditions.
Verify physical particle states, voltage response, and lithium conservation before testing plating kinetics.
The DFN/P2D model cannot produce metallic lithium.
Add a plating/stripping reaction coupled to electrolyte concentration and the local solid–electrolyte potential difference.
Retained. The reaction becomes active during fast charging and produces a spatial deposition rate.
The reversible plated-lithium state can decrease during stripping.
Compare instantaneous current, net plated lithium, and a deposition-only accumulated state.
Retain the deposition-only state for the onset calculation.
The experimental target is a capacity-normalized accumulated quantity.
Define a constant-fraction representation as Qirr=0.30Qplate and apply a threshold equal to 0.05% of areal capacity.
Retained as the reference representation. Capacity normalization gives thresholds of 37.8 and 55.8 C m -2 for A1 and A2.
Appendix
Table 5: EvoSim’s physical hypotheses and decisions in the lithium-plating case.
Parameter
Retained value
Graphite thickness ( μ m)
47/70
Electrolyte volume fraction
0.374/0.354
Active-solid volume fraction
0.5954/0.5901
Graphite particle radius ( μ m)
4
Initial electrolyte concentration (mol m -3 )
1200
i0,Liref (A m -2 )
0.75
Appendix
Table 6: Selected quantities in the retained rate-dependent lithium-plating model. Slash-separated entries give A1/A2 values.
Elec.
Temp. ( ∘ C)
Rate
Data role
Exp. (%)
Constant (%)
Rate dep. (%)
Constant error (pp)
Rate-dep. error (pp)
A1
25
4 C
Validation
61.44
65.31
62.47
+3.88
+1.04
A1
25
5 C
Training
57.33
55.94
55.07
-1.39
-2.26
A1
25
6 C
Training
48.25
45.47
47.83
-2.78
-0.43
A1
35
4 C
Training
67.11
69.81
66.96
+2.69
-0.15
A1
35
5 C
Validation
59.28
61.77
61.29
+2.49
+2.01
A1
35
6 C
Training
54.41
54.02
56.06
-0.39
+1.65
Appendix
Table 7: Pointwise comparison of experimental onset SOC with the constant-fraction and rate-dependent representations. The fixed split contains 14 training conditions and six validation conditions. Model selection uses only the training conditions. Errors are prediction minus experiment in SOC percentage points.
Condition
Experiment (%)
Published model (%)
EvoSim (%)
Published error (pp)
EvoSim error (pp)
A1, 25 ∘ C, 4C
61.44
60.53
62.47
-0.90
+1.04
A1, 35 ∘ C, 5C
59.28
59.58
61.29
+0.30
+2.01
A1, 45 ∘ C, 6C
56.75
60.11
60.45
+3.36
+3.69
A2, 25 ∘ C, 2C
70.39
69.47
69.94
-0.92
-0.45
A2, 35 ∘ C, 4C
49.80
47.22
52.40
-2.58
+2.60
A2, 45 ∘ C, 6C
46.49
37.59
45.54
-8.90
-0.95
Appendix
Table 8: Comparison on the six conditions reserved for EvoSim validation. EvoSim does not use these conditions for parameter identification. The source-study model used the full 20-condition matrix during model development ( Konz et al., 2023 ) . Errors are prediction minus experiment in SOC percentage points.
Candidate and observation
Model definition or modification set
Reported RMSE
Supported conclusion
Starting diagnostic
Four electrode filling states; equilibrium potentials; charge-transfer polarization; internal resistance
32.74 mV
The main voltage level is present. The response after current changes remains incomplete.
Electrochemical foundation
Four electrode filling states with equilibrium, charge-transfer, and ohmic terms.
32.12 mV
Establishes the starting physical voltage model.
Extended physics-based model
Solve electrolyte transport, solid diffusion, Faradaic intercalation, double-layer charging, and temperature-dependent transport and kinetics together.
Run result
The complete voltage trace remains finite and satisfies the physical checks.
Held-out runs
Apply the fixed physics-based model to complete traces at temperatures excluded from parameter identification.
7.62 and 8.97 mV
These are results for individual runs at 30 ∘ C DST and 20 ∘ C US06.
Appendix
Table 9: Recorded hypotheses, model changes, and decisions in the dynamic-voltage model sequence.
Figure 5: Selected complete-trace runs of the physics-based voltage model. Figure 3 c reports two separate held-out results.
Figure 6: Remaining complete-trace fits using all available data. Together with Figure 5 , these panels contain all 16 traces.
Model structure
Comparison condition
RMSE (mV)
Same A123 trace and inputs; original identification procedures
Physics-based model (this work)
Full-data run; not held out
6.08
CPG-SPMT ( Guo and Couto, 2026a )
Published result on the same condition
32.10
DST/US06 dynamic-voltage results reported by each study
Physics-based model (this work)
Held-out results shown in Figure 3
7.62, 8.97
NDCTNet ( Tu et al., 2024 )
Reported results
12.25, 13.01
Appendix
Table 10: Reported dynamic-voltage RMSE under DST or US06 operating conditions. The first block uses the same A123 A1-007 25 ∘ C DST measurement trace and inputs.
Question
Evidence
Conclusion supported
Does Scientific Evolution produce a usable procedure?
An earlier continuous-state model gives 88.518 mV error against the reference implementation. EvoSim diagnoses the discrete state update, checks the corrected implementation at 4.1136 mV, and saves the procedure.
EvoSim converts its own model evidence into a checked action with stated conditions.
Does the module improve the next decision?
Baseline, Naive Memory, and Scientific Evolution use the same comparison task, starting model, data, solver, checks, and maximum of 12 attempts. The first three attempts are the matched comparison. Naive Memory and Scientific Evolution use the same source investigation.
Scientific Evolution reaches 22.015 mV. The chronological summary remains at 89.383 mV.
Does recursive updating occur?
Four successive research steps connect failure diagnosis, a checked correction, a 24-condition application, and the controlled comparison.
Each of the first three results changes the procedure used in the following investigation. The following result supplies evidence for the next update.
Does the updated procedure help later work?
The procedure changes the later state implementation and reaches a valid lower-error model on attempt 2. Three related-task records report fewer revisions or lower error after later use.
The updates change later model decisions, implementation effort, and numerical results within and across tasks.
Does the effect persist across repeated tasks?
Forty-five paired experiments compare Baseline with Scientific Evolution after the procedures and pair list are fixed.
Model / physics failures decrease from 33 to 21; the exact paired test remains significant after correction across six labels.
Appendix
Table 11: Evidence used to evaluate Scientific Evolution.
Condition
Earlier record
Information available for the next decision
Naive Memory
“A later version represented the sample-wise state recurrence explicitly and reduced the 30 degC DST COMSOL-to-Python voltage RMSE to about 4 mV.”
Identifies the general cause and records the earlier result. It does not state the recurrence, sample timing, state binding, initial-state check, temperature rule, or invalid-output conditions.
Scientific Evolution
“For every k>0 , use current and temperature from sample k ” and “Export time, current, temperature, all four COMSOL dependent states, and terminal voltage directly from COMSOL.”
Gives the exact recurrence, initial values, one temperature conversion, four dependent-state equations, direct-output checks, supported conditions, and failure conditions.
Appendix
Table 12: EvoSim-generated records available to the two conditions. Both records come from the same earlier investigation.
Condition
Earlier record available to the run
Experimental-voltage RMSE (mV)
Baseline
No retained record from the earlier investigation
89.478 mV
Naive Memory
EvoSim-generated chronological summary of the same earlier investigation
89.383 mV
Scientific Evolution
EvoSim-generated procedure produced and checked from the same earlier investigation
22.015 mV
Appendix
Table 13: Best valid result produced within the first three attempts (physics-based model–experimental-voltage RMSE).
Budget level
Baseline
Naive Memory
Scientific Evolution
Interpretation
RMSE (mV)
B=3
89.478
89.383
22.015
First three submissions
B=6
89.478
89.383
–
Baseline and Naive Memory persistence check
B=12
89.478
89.383
–
Baseline and Naive Memory persistence check
Appendix
Table 14: Best valid RMSE in the first three submissions and during the Baseline and Naive Memory persistence check.
Attempt
Baseline
Naive Memory
Scientific Evolution
1
Continuous physics-based model state equations (89.478; valid)
Continuous Global Equations starting model (89.383; valid)
Setup failed; no model output
2
Previous-sample current and temperature (89.666; valid)
Discrete states and event update; build failure
Exact ZOH, one temperature conversion, and four algebraic states (22.015; valid)
3
Exact ZOH interpolation without the required dependent-state representation (83.638; invalid)
Event reinitialization leaves states at their initial values (394.480; valid)
Nearest/sample-hold state transfer; diagnostic export, not retained
Appendix
Table 15: First three modeling attempts in the three-condition comparison.
Attempt
Baseline
Naive Memory
4
Solver bypass: bypasses the required model equations and returns a supplied target output (invalid)
Global event selection activates updates, but the surface state collapses to the average state (89.578; valid)
Exact matrix recurrence, average state first (89.757; valid)
8
Repeats the same solver bypass (invalid)
Surface-state update created first (89.757; valid)
9
Repeats the same solver bypass (invalid)
Shifted zero-order-hold current (89.573; valid)
Appendix
Table 16: Baseline and Naive Memory attempts after the first three attempts. A candidate is invalid when it bypasses the required solver-based state and voltage equations.
Task
Fixed task and acceptance check
Observed outcome
What the outcome demonstrates
Voltage-model implementation
Four required physics-based model states, fixed terminal-voltage relation, complete finite export
The checked procedure changes the implementation and lowers the voltage error
Capacity matching
Specified cell configurations; every candidate must build, solve, and export
Required attempts decrease from 8 to 2
A retained procedure reduces repeated model revisions
P2D–thermal coupling
Coupled model must solve and export the complete temperature trace
Model changes decrease from 10 to 1
A retained procedure changes how the coupling investigation is organized
Thermal-mechanism selection
Particle size, polarization, boundary heat, and convection use one temperature metric
Temperature RMSE decreases from 9.30 to 3.74 ∘ C
A common-check comparison changes both candidate selection and prediction error
Appendix
Table 17: Outcomes of the four task-level model-development comparisons.
Earlier investigation
Later investigation
Reused modeling decision
Result in the later task
Capacity investigation
Specified-configuration capacity-matching task
Apply the tested capacity-matching order and stop only after the configuration, solve, and export checks agree
Required attempts decrease from 8 to 2
Thermal-mechanism investigation
Subsequent P2D–thermal coupling task
Compare independent coupling mechanisms before repeatedly modifying one candidate; retain only a candidate that passes the common temperature check
Sequential model changes decrease from 10 to 1
Thermal-mechanism investigation
Thermal-mechanism selection target
Use one temperature metric and compare particle-size, polarization, boundary-heat, and convection candidates under the same starting conditions
Temperature RMSE decreases from 9.30 to 3.74 ∘ C
Appendix
Table 18: Use of saved procedures in three related modeling tasks. Each procedure is completed before the later task begins. Each later task uses its own acceptance check.
Research step
New evidence
Change to the saved procedure
Observed later use
Failed earlier model
Continuous state equations solve but give 88.518 mV RMSE against the official discrete reference implementation. The surface states contain the largest differences.
Record the discrete state update as the cause to test before changing parameters or the voltage equation.
The next model implements the exact sample-wise zero-order-hold update.
Single-condition correction
The corrected 30 ∘ C DST model gives 4.1136 mV RMSE against the reference and passes the state and output checks.
Retain the state update, temperature conversion, four-state representation, and output checks as a reusable procedure.
A later investigation applies the procedure to every available temperature and drive-cycle condition.
Twenty-four-condition application
All 24 conditions solve. Their mean RMSE against the reference implementation is 4.4625 mV.
Expand the supporting evidence from one trace to 24 conditions while retaining the one-second CPG-SPMT equations and input rules as the stated scope.
The checked procedure becomes available to the later controlled voltage-modeling task.
Controlled comparison
The first application detects a repeated temperature conversion. The next attempt corrects the conversion, represents all four dependent states, and gives 22.015 mV RMSE against measured voltage.
Add the failed double conversion as direct evidence for the temperature check and record the state-output check with the valid implementation.
The procedure changes the modeling sequence and reaches the valid low-error model on attempt 2.
Appendix
Table 19: Successive evidence and updates for the exact zero-order-hold procedure.
Record element
Recorded evidence
Consequence for the next decision
Earlier candidate
The earlier voltage investigation used continuous state equations in the physics-based model. The candidate passed the finite-output and input-value checks but failed the target state and voltage comparison.
The discrepancy motivates separate decisions about the model and the state-update procedure.
Model decision
Reject the continuous implementation for the stated task.
Use the diagnosed discrepancy to examine the state-update procedure.
Proposed procedure
The logs support an exact sample-wise ZOH update. Convert the trace temperature once. Use nearest-neighbor current and temperature inputs with the recorded half-step timing. Represent the computed states as linear time functions on the one-second grid. Preserve the stated equations, voltage bounds, and checked solver settings.
Define a specific operation and the checks needed to evaluate it.
Procedure decision
Retain the exact ZOH procedure for the one-second, 30 ∘ C DST task after the implementation checks pass.
Make the procedure available with the same unit, state-variable, and output checks.
Saved procedure
EvoSim stores the procedure, its source evidence, supported conditions, and file list with the run.
A later attempt can retrieve the operation and its conditions from one record.
First later use
The first application converts a ∘ C value to Kelvin twice and fails the submission check. The next attempt uses one conversion, four algebraic states in the physics-based model, and the fixed output format.
Correct the temperature conversion before testing the retained operation.
Appendix
Table 20: Model and procedure decisions in the fixed four-state voltage task.
Quantity
Count
Modeling tasks with repeated runs
11
Baseline runs
45
Scientific Evolution runs
45
Unique runs
90
Baseline–Scientific Evolution pairs
45
Appendix
Table 21: Run accounting for the paired-run analysis.
Failure mode
Recorded condition
False completion
Completion is declared while a required output, check, or scientific result still fails
Tool operation
Model inspection, build, save, export, or data access is executed incorrectly
Model / physics
A required physical process is absent or inactive, or the tested model form cannot reproduce the target behavior under the task criterion
Data / protocol
Inputs, units, initial conditions, stopping rules, time alignment, or data roles are bound incorrectly
Numerical solver
The solve fails, diverges, becomes singular, or produces non-finite values
Looping / stagnation
Repeated actions or failed recovery do not advance the task to a supported decision
Appendix
Table 22: Operational definitions used for failure-mode labels.
Failure mode
Baseline count
Scientific Evolution count
Baseline- only
Both
Evolution- only
Net decrease (%)
False completion
12
0
12
0
0
100.0%
Tool operation
19
2
17
2
0
89.5%
Model / physics
33
21
15
18
3
36.4%
Data / protocol
14
0
14
0
0
100.0%
Numerical solver
5
2
5
0
2
60.0%
Looping / stagnation
25
5
21
4
1
80.0%
Appendix
Table 23: Failure-mode transitions across 45 same-task pairs.
Failure mode
Baseline- only ( b )
Evolution- only ( c )
Exact p -value
Holm-adjusted p -value
False completion
12
0
0.000488
0.001465
Tool operation
17
0
0.000015
0.000076
Model / physics
15
3
0.007538
0.015076
Data / protocol
14
0
0.000122
0.000488
Numerical solver
5
2
0.453125
0.453125
Looping / stagnation
21
1
0.000011
0.000066
Appendix
Table 24: Exact paired tests for the failure-mode transitions.
Failure mode
Baseline tasks
Scientific Evolution tasks
Baseline runs
Scientific Evolution runs
False completion
1
0
12
0
Data / protocol
4
0
14
0
Tool operation
4
1
19
2
Looping / stagnation
6
2
25
5
Numerical solver
2
1
5
2
Model / physics
7
7
33
21
Appendix
Table 25: Number of tasks and runs with each failure mode. A task is counted when the failure appears in at least one run. Run counts include repeated experiments on the same task.
Measure
Baseline total
Scientific Evolution total
Ratio (%)
Final-check failures
167
53
31.7
Unclosed work items
548
187
34.1
Active runtime (h)
1086.3
623.6
57.4
Elapsed time (h)
1217.7
708.7
58.2
Assigned work items
3792
3169
83.6
Appendix
Table 26: Process totals and ratios across the 45 same-task pairs.
Task family
Scientific objective
Required model output
Evidence required for completion
Parameter identification and calibration
Identify parameter values that explain the recorded observations
Transport, kinetic, geometric, or thermal parameters with specified bounds
Successful solve, exported observables, and comparison with the recorded observations
Forward prediction
Predict a new protocol or operating condition with a fixed model
Voltage, state of charge, capacity, or lifetime curve or map
Prediction metric on the specified evaluation protocol
Mechanism and multiphysics
Determine whether an added physical process is needed and correctly coupled
Executable mechanism, state, equation, or multiphysics coupling
Implementation check, solution check, and comparison with observations
Model diagnosis and procedure
Diagnose a failed model change and establish a reproducible correction
Failure diagnosis, corrected procedure, or verified negative result
Build, solve, and export checks with a record that links the cause to the correction
Appendix
Table 27: Task definitions and completion requirements in the recorded run collection.
Quantity
Count
Recorded runs
179
Runs for application tasks
64
Runs for physics-based battery tasks
115
Task families in the recorded run collection
32
Task families in the cross-task analysis set
26
Run-level records in the cross-task analysis set
65
Appendix
Table 28: Counts for runs, task families, and evaluations.
Outcome
Task families
Count
Reason
Complete evaluation
Molicel P42A transfer; LG M50T diagnostic-to-another-condition; degradation-mode prediction; Panasonic 18650PF drive-cycle evaluation; pouch-cell thermal prediction; LG M50-family cross-dataset transfer; COBRAPRO UDDS; CPG-SPMT FUDS; SEI mechanism competition
9
Fixed candidate, comparison target, numerical result, final decision, and supported conditions are available.
Comparison used for later model changes
Panasonic multi-temperature model; LG M50 full-matrix model
2
The comparison later enters model development, or the final conclusion uses a metric from an earlier model.
Development record
Lithium-plating temperature–rate boundary; lithium-plating risk over temperature and rate
2
The result contributes to threshold construction or has no matching metallic-lithium outcome.
Automated simulator construction requires distributional fidelity, distinguishing it from generic code generation. We identify two failure modes in long-horizon LLM agents: contextual drift and optimization instability arising from conflating structural and parametric errors. We propose SOCIA-EVO, a dual-anchored evolutionary framework. SOCIA-EVO introduces: (1) a static blueprint to enforce empirical constraints; (2) a bi-level optimization to decouple structural refinement from parameter calibration; and (3) a self-curating Strategy Playbook that manages remedial hypotheses via Bayesian-weighted retrieval. By falsifying ineffective strategies through execution feedback, SOCIA-EVO achieves robust convergence, generating simulators that are statistically consistent with observational data. The code and data of SOCIA-EVO are available here: https://github.com/cruiseresearchgroup/SOCIA/tree/evo.
Yuncheng Hua, Sion Weatherhead, Mehdi Jafari +2
School of Computer Science and Engineering, University of New South Wales, Australia
Astra can act, yet reliable manipulation depends on the system through which it observes and controls the world. We introduce PhysEvo, a framework for physical recursive self-improvement (RSI) around a single frozen model. A task agent executes robot tasks; a meta-agent uses the resulting trajectories to diagnose failures, revise tools and skills, and test corrections. The meta-agent can also improve its own diagnostic tools, so retained revisions support both later action and later self-improvement. This process develops joint-level control, evidence-seeking observation, and reusable manipulation skills without model-weight updates or a separately trained action policy. Across 42 RoboDojo tasks, held-out-layout evaluation of retained task-specific deployment versions yields a five-dimension average score of 68.14/100 and 62.00% success, compared with 47.17% for RoboDawn's one-shot Astra agent, the strongest published reference in our comparison. On eight manipulation tasks challenging direct Astra, PhysEvo achieves 55.00% success, compared with 1.25% for the direct-Astra reference. Deploying the simulation-evolved harness on AgileX PiPER and continuing skill revision yields 90.60/100 average score and 84.00% success across 25 trials on five real-world tasks. PhysEvo turns the consequences of action into persistent, testable changes to how a frozen model acts and improves.
Wenqing Tian, Zeyu Zhang, Zhaocheng Liu +3
University of Chinese Academy of Sciences · Institute of Automation, Chinese Academy of Sciences · Independent Researcher +1
The increasing adoption of Large Language Models (LLMs) has enabled AI scientists to perform complex end-to-end scientific discovery tasks requiring coordination of specialized roles, including idea generation and experimental execution. However, most state-of-the-art AI scientist systems rely on static, hand-designed pipelines and fail to adapt based on accumulated interaction histories. As a result, these systems overlook promising research directions, repeat failed experiments, and pursue infeasible ideas. To address this, we introduce EvoScientist, an evolving multi-agent AI scientist framework that continuously improves research strategies through persistent memory and self-evolution. EvoScientist comprises three specialized agents: a Researcher Agent (RA) for scientific idea generation, an Engineer Agent (EA) for experiment implementation and execution, and an Evolution Manager Agent (EMA) that distills insights from prior interactions into reusable knowledge. EvoScientist contains two persistent memory modules: (i) an ideation memory, which summarizes feasible research directions from top-ranked ideas while recording previously unsuccessful directions; and (ii) an experimentation memory, which captures effective data processing and model training strategies derived from code search trajectories and best-performing implementations. These modules enable the RA and EA to retrieve relevant prior strategies, improving idea quality and code execution success rates over time. Experiments show that EvoScientist outperforms 7 open-source and commercial state-of-the-art systems in scientific idea generation, achieving higher novelty, feasibility, relevance, and clarity via automatic and human evaluation. EvoScientist also substantially improves code execution success rates through multi-agent evolution, demonstrating persistent memory's effectiveness for end-to-end scientific discovery.