Research

Battery safety

Goal. Predict how a lithium-ion or solid-state cell fails: lithium-dendrite growth, internal short circuit, heat generation and thermal runaway.

Significance. Each step is difficult to observe directly, and abuse tests are slow, costly and hazardous. Coupled phase-field and multiphysics simulation resolves the full sequence; machine-learning surrogates reduce it to a fast early-warning estimate; and AI tools help decide which experiments to run.

DendriX: 3D dendrite growth
Charge analysis and AI surrogate
From internal short to heat
ThermalRunawayAI: literature assistant
AI-guided lab workflow

DendriX: 3D dendrite growthDendriX is a phase-field platform for lithium-dendrite growth, designed for extension to polycrystalline solid
electrolytes, electronic leakage and heat generation. It runs in 2D and in 3D on HPC clusters, which allows a
dendrite to be followed through a realistic grain structure from the anode to the opposite electrode.
Lithium dendrite (red) growing through a 3D polycrystalline solid electrolyte on a 256³ grid. in preparationGeneration of the electrolyte microstructure: about 850 grains nucleate and fill the domain before dendrite growth starts. in preparationRotating view of a single 3D dendrite in the polycrystal. in preparation2D simulation: dendrites (yellow) advancing through a polycrystalline electrolyte. in preparationSpike-type (a) and tree-type (c) 3D dendrites over time, with cross sections (b, d). W. Yang, R. Wang, et al., SoftwareX 2026, Fig. 4.Multiphysics coupling: dendrite growth in a polycrystal (a–d) and the corresponding cell voltage without (e) and with (f) the dendrite. SoftwareX 2026, Fig. 5.
The physicsA phase-field order parameter distinguishes lithium metal from electrolyte. The interface evolves by Allen–Cahn
kinetics driven by the Butler–Volmer reaction, coupled with Li+ transport and the electric potential.
In a solid electrolyte, grain boundaries carry their own transport and reaction properties.
Papers: SoftwareX 2026
DendriX: team code (W. Yang, R. Wang, et al., L. Chen lab, U. Michigan); R. Wang is a co-developer. 3D solid-electrolyte (polycrystal) simulations: R. Wang. Paper figures reused under CC BY-NC 4.0.
Charge analysis and AI surrogatein preparationBefore a dendrite bridges the electrodes, electronic leakage through the electrolyte can produce a soft short.
The dendrite microstructure is coupled to a charge-transport model to map the leakage current, and a neural network
trained on thousands of these simulations reproduces the result in a fraction of a second, which makes it suitable
for early warning.
Own code In-house solver and surrogate, written from scratch in Python (PyTorch); no commercial software.
Left: dendrite morphology (black) in an LLZO solid electrolyte. Right: computed leakage-current density on the same grid (log scale).Computed net ionic charge around the dendrite network.Training set: about 2,300 phase-field dendrite morphologies, split into training and validation sets for a convolutional neural network.Surrogate prediction of cell voltage against the full simulation on held-out cases.Wall-clock time per case: 1.78 h for the full simulation, 310 ms for the surrogate (about 20,000× faster).
The physicsFor a given dendrite morphology, a steady charge-transport problem gives the potential and current density:
∇·(σ ∇φ) = 0,  σ = σion + σe
The electronic conductivity σe is computed from defect chemistry, with separate properties for the
grain interiors and the grain boundaries.
Model, code and surrogate: R. Wang, L. Chen lab, U. Michigan. Manuscript in preparation.
From internal short to heatin preparationThe dendrite model is coupled to heat generation and a cell-level thermal model, so that the temperature response
of the cell can be followed from the moment an internal short forms.
DendriX simulation: as the dendrite network connects the electrodes, Joule heat and cell temperature rise.
The physicsThe short-circuit current produces Joule heat, q = σ|∇φ|². The temperature follows
ρcp ∂T/∂t = ∇·(k∇T) + qJoule + qreaction
where the reaction heat comes from Arrhenius-type decomposition reactions that activate as the cell heats up.
Models: R. Wang (DendriX and COMSOL). The COMSOL cell model builds on the internal-short model of J. Su. Manuscript in preparation.
ThermalRunawayAI: literature assistantin preparationThermal-runaway data are scattered across many papers that use different cells, tests and units.
ThermalRunawayAI is an agentic retrieval-augmented (RAG) assistant over a curated thermal-runaway literature. It
answers with citations to the specific paper, figure or table, and reports when the collection does not
contain the answer.
Pipeline: about 60 papers are split into text, figures and tables, searched with hybrid retrieval and re-ranking, and answered with citations.Example answer with citations: ranking of insulation materials for blocking propagation, linked to the source paper.When a value is not reported in the collection, the assistant states that it is missing.
Project lead: R. Wang. Developed with high-school researchers William Li (retrieval engine) and Jonathan Mei (PDF text and figure extraction). Paper collection: J. Su; L. Chen lab. Paper in preparation.
AI-guided lab workflowin preparationBattery abuse samples are expensive and the tests are hazardous, so the number of experiments should be kept low.
The workflow connects literature mining by AI, model-guided experiment design, and testing (with a robotic
setup under construction) in one loop. For daily modeling, AI agents prepare and check simulation runs on a
licensed workstation and the Great Lakes cluster, with each next step approved by the researcher.
One loop of the workflow: literature mining supplies model parameters, the model guides experiment design, and test outcomes feed back.
Workflow design: R. Wang, L. Chen battery lab, U. Michigan. Experimental data: J. Su. Robotic setup under construction.