EvoSim: AI That Evolves Physics Models from Experimental Discrepancies

From Parameter Tuning to Equation Revision

AI excels at formal reasoning tasks such as mathematics and programming, but struggles with the microscopic physical world underlying materials R&D and device design. Physical models bridge observable phenomena and internal mechanisms by expressing transport, reaction, and energy exchange as state variables and equations. However, real systems couple multiple physics fields, and key internal states are often unmeasurable. When a model disagrees with experiment, an AI must diagnose whether the error stems from inaccurate parameters, numerical issues, or missing physics — and then reuse that diagnostic skill for future tasks.

To address this, Professor Qiang Zhang's group at Tsinghua University, together with Qingyu Technology, the University of Hong Kong, and the Beijing Huairou Laboratory, developed EvoSim . Its core advance is shifting AI from "parameter tuning under fixed equations" to "mechanism tracing, equation revision, and modeling-method improvement" driven by experimental discrepancies. A single trial-and-error cycle simultaneously yields a more accurate model and a reusable modeling capability. The system is validated on two industrial-grade battery scenarios.

Case Study 1: Lithium-Plating Onset in Fast Charging

Fast charging pushes batteries toward lithium plating, where lithium deposits as metal instead of intercalating into graphite, consuming cyclable lithium and threatening safety. Predicting the onset state-of-charge (SOC) requires accounting for temperature, electrode design, and protocol — many of which are not directly measurable online. EvoSim starts from a base electrochemical model augmented with plating/stripping reactions. Systematic deviations across rates pointed to a deeper assumption: that deposited lithium converts to irreversible loss at a fixed fraction. EvoSim rewrote this relationship as rate-dependent, letting the equations reflect the fast-charging timescale.

Experiments used 61 test cells covering two graphite electrode designs, 25–45 °C, and multiple charge rates, providing 20 plating-onset conditions. On the test set, EvoSim's model achieved a mean absolute error of 1.79% SOC , outperforming the best historical human-expert models on the same data.

Case Study 2: Dynamic Voltage Prediction Under Real Driving Loads

Dynamic voltage trajectories determine whether a battery can deliver power during acceleration, accept regenerative braking energy, and stay within operating limits — critical for battery management systems. Real driving currents change frequently, while electrode states, electrolyte transport, and interfacial charging respond at different speeds. EvoSim reconstructed the electrode state-update scheme and incorporated these coupled processes into the solver.

Validation covered two drive cycles (DST and US06) at eight temperatures, totaling 16 full trajectories and 109,746 second-level samples . At 30 °C DST, EvoSim's root-mean-square voltage error was 7.62 mV ; at 20 °C US06, 8.97 mV . Both are lower than the best previously published human-expert cell models on the corresponding datasets.

Figure 2 (from the paper) shows lithium-plating onset prediction and full voltage trajectories under dynamic loads.

Self-Evolution Loop: Models and Methods Co-Improve

The dynamic voltage task illustrates the reciprocal evolution. Early models converged numerically but produced internal electrode states inconsistent with reference calculations. Instead of merely adjusting parameters, EvoSim traced the discretization of state updates. Each failure or success updated the system's checklist and decision order for subsequent modeling rounds — a recursive self-improvement.

To quantify the accumulated method knowledge, the team ran three controlled groups on the same dynamic voltage task with a fixed submission budget:

Group 1: full history retained

Group 2: only attempt summaries retained

Group 3: EvoSim (full self-evolution)

EvoSim's best valid error was 75.4% lower than Group 1, while Group 2 matched Group 1. The gap shows that quality gains come from learning "how to proceed next" from model outcomes, not merely remembering past attempts.

This feedback extends to related tasks. In electro-thermal coupling modeling, the system switched to parallel mechanism exploration followed by model revision, reducing consecutive model modifications from 10 to 1 . Across 45 runs covering 11 tasks , model-construction or mechanism-failure occurrences dropped 36.4% and cumulative runtime fell 42.6% .

Toward a General Scientific Research Paradigm

EvoSim demonstrates a pathway for language models to continuously understand the physical world: experiments flag where physics models fail; the language model revises the model and offers new explanations; the revision process in turn improves its research methodology. Physical models and scientific capability co-evolve on the same evidence chain.

Battery tasks provide a quantifiable starting point. Next steps include selecting experimental systems in materials science, chemical engineering, and bioengineering to test whether this co-evolution generalizes. Along this trajectory, AI's grasp of the physical world is measured not only by single predictions but by its ability to propose better models and better research methods from every collision with reality.

Paper: EvoSim: Learning to Model, Modeling to Learn —

https://arxiv.org/abs/2610.11344