Battery materials discovery is one of the killer applications for machine learning potentials. You need to screen hundreds of compositions, compute migration barriers in supercells with thousands of atoms, and run long molecular dynamics trajectories to measure ionic conductivity. All three of these used to be bottlenecks because DFT is too slow. With MACE-MP-0, they are practical for a single graduate student. This post walks through how to use it.
What problems battery R&D needs to solve
A working battery has four components that materials scientists obsess over:
- Cathode. The positive electrode. Stores lithium at high voltage. Examples: LCO, LFP, NMC, LMO.
- Anode. The negative electrode. Stores lithium at low voltage. Examples: graphite, silicon, lithium metal.
- Electrolyte. The medium that lets lithium ions move between the electrodes. Can be liquid or solid.
- Interphases. The solid electrolyte interphase on the anode and the cathode electrolyte interphase on the cathode. These form in situ and are critical for cell lifetime.
Each of these has different modeling needs. Cathodes want formation energies, voltage profiles, and migration barriers. Electrolytes want long MD trajectories to measure ionic conductivity. Interphases want reactive chemistry.
Cathode screening
The classic cathode screening question is: given a crystal structure and a candidate composition, what is the average voltage versus lithium metal, and what is the lithium migration barrier?
Voltage profiles
Voltage is computed from formation energy differences between the fully lithiated and fully delithiated structures. With MACE-MP-0, a 100-atom supercell relaxation finishes in seconds instead of hours. You can screen 100 compositions overnight on a laptop. A GPU makes it an afternoon job.
Migration barriers
Lithium migration barriers determine rate performance. You compute them with nudged elastic band calculations between neighboring lithium sites in the cathode lattice. With DFT, a single NEB calculation takes hours. With MACE, it takes minutes. That turns migration barrier measurement from a rare, careful endeavor into a routine screen.
Solid electrolytes
Solid electrolytes are the holy grail of safer, higher-energy batteries. The screening question here is usually ionic conductivity. You set up a supercell, run a long molecular dynamics trajectory at elevated temperature, and measure the mean squared displacement of lithium atoms as a function of time.
The challenge is that ionic conductivity requires nanosecond-scale trajectories at multiple temperatures, and DFT-based MD is limited to picoseconds. MACE makes nanosecond trajectories feasible, which finally lets you compute reliable conductivities from first principles.
Common solid electrolyte chemistries like Li10GeP2S12, Li7La3Zr2O12 garnet, and argyrodite Li6PS5Cl are all well represented in the MACE-MP-0 training distribution. You can run them without fine-tuning for a first pass, then refine with a small custom dataset if needed.
A practical workflow
Here is a workflow that a small group can run on a handful of laptops plus occasional GPU access:
- Day 1: Pull 100 candidate compositions from the Materials Project, filter by stoichiometry and stability, and relax each with MACE-MP-0.
- Day 2: Compute voltages for the top 30 by relaxing fully delithiated counterparts.
- Day 3: Run NEB migration barriers on the top 15.
- Day 4-5: For the top 5 candidates, run 1 nanosecond MD at 600 K and 900 K, compute diffusion coefficients, and extrapolate room-temperature conductivity via Arrhenius.
- Day 6: Spot-check the top candidate with DFT single-point energies on representative frames to validate the MACE results.
- Day 7: Write up and send the top 3 candidates to the synthesis team.
A week of work at laptop scale replaces what used to be months on a cluster. The ratio is not subtle.
When to fine-tune
The foundation checkpoint is a great starting point, but for battery work you often want higher accuracy than the default. Signs you should fine-tune:
- Spot-check DFT energies disagree with MACE by more than about 50 millielectronvolts per atom on your target chemistry.
- Migration barriers from MACE are more than 100 millielectronvolts off the DFT reference.
- Your composition includes unusual combinations of elements that are sparse in the training distribution.
A useful fine-tuning set is typically 200 to 1000 DFT single-points covering relaxed structures, NEB images, and sampled MD frames from your chemistry. The MACE training framework supports this directly.
What MACE cannot do well (yet)
MLIPs are not a universal solvent. For battery work specifically, a few things are still hard:
- Electronic structure details. Charge transfer, redox state changes, and magnetism in transition metal oxides can be subtle. MACE predicts the aggregate energies correctly but does not give you the electronic picture.
- Reactive SEI chemistry. Reactions between electrolyte molecules and electrode surfaces are typically outside the training distribution and need DFT.
- Very high voltage edge cases. Some high-voltage cathodes involve oxidation states that are rare in the training data.
For these, use DFT on the specific configurations you care about and use MACE for the rest of the screening.
Bottom line
Battery materials discovery is exactly the kind of problem that MLIPs were built for: lots of similar structures, long trajectories, many candidates, limited budget. MACE-MP-0 is accurate enough for a first pass on most common chemistries and fine-tunes well for the edge cases. If you are running a battery materials group in 2026 and still doing all your screening with pure DFT, you are leaving a lot of speed on the table.