The battery bottleneck is measurement — and new tools are attacking it
A tabular model predicts electrolyte conductivity from 40 labeled samples. A solid-electrolyte study says the standard high-symmetry model gets the energetics wrong.
- TabPFN achieves a 5.56% MAPE on the AMPERE conductivity dataset using 40 labeled samples. Self-reported m:…3376
- Bayesian optimization using TabPFN reduces experimental evaluation costs by 45.6% compared to random search. Self-reported m:…3376
- TabPFN provides a direct objective surrogate for electrolyte formulation search, requiring 38.2% fewer evaluations than random search. Self-reported m:…3376
- Using ideal high-symmetry references for LPSC leads to qualitative errors in predicting energetics for In/O co-doping and S/O compositional mixing. Self-reported m:…9446
- GITTcalc automates the extraction of lithium diffusion coefficients from GITT experimental data and includes sanity checks to verify if fundamental GITT assumptions are respected. Self-reported m:…3375
Few samples, and a usable prediction
A ChemRxiv preprint reports that TabPFN, a tabular foundation model, achieves a 5.56% MAPE on the AMPERE conductivity dataset using 40 labeled samples. Electrolyte conductivity is a property battery researchers measure to judge whether a formulation will carry charge well, and the point of the result is how little labeled data the model needed to predict it. m:…3376
The same paper reports that Bayesian optimization using TabPFN reduces experimental evaluation costs by 45.6% compared with random search, and that the model requires 38.2% fewer evaluations than random search. The model serves as a direct objective surrogate for electrolyte formulation search, which means it stands in for the lab measurement while the search narrows down candidates, and the lab only tests the survivors. m:…3376
The combination matters because electrolyte formulation is a search problem: the space of possible mixtures is vast, and each lab test costs time and money. A surrogate that is accurate after 40 labeled samples, and that cuts the evaluations needed by 38.2%, turns a search that was too expensive to run into a search a small team can afford. m:…3376
The standard model gets the energetics wrong
Another preprint, on argyrodite-type solid electrolytes with the formula Li6PS5X, reports that structural relaxation lowers the energy of these electrolytes compared with ideal high-symmetry models. The distorted configuration, called d-LPSX, represents a dynamically stable local minimum, which means the lower-symmetry shape is a real resting state of the material rather than an artifact of the calculation. m:…9446
The practical consequence is that using ideal high-symmetry references for LPSC leads to qualitative errors in predicting energetics for In/O co-doping and S/O compositional mixing. In plain terms, a model that assumes the tidy symmetric structure will give wrong answers about which dopants and mixtures are energetically favorable, and the error is qualitative, not a small rounding difference. m:…9446
This is a warning aimed at the modelling layer itself. If the reference structure is wrong, every downstream prediction built on it inherits the error, and no amount of compute spent on the wrong reference will fix it. The fix the paper points to is to relax the structure and model the distorted configuration that the material actually adopts. m:…9446
Automating the measurement itself
A further preprint automates a measurement rather than a prediction. GITTcalc automates the extraction of lithium diffusion coefficients from GITT experimental data, a standard technique for measuring how fast lithium moves through a battery material. The tool includes sanity checks to verify whether fundamental GITT assumptions are respected during data analysis, and it provides interoperability with OriginLab for battery data processing. m:…3375
The sanity checks are the interesting part. Diffusion coefficients extracted from data that violate the technique's assumptions are worse than useless, because they look like measurements while being artifacts. A tool that flags the violation before the number enters a paper or a model is doing quality control at the point of measurement, which is where bad numbers are cheapest to catch. m:…3375
Why it matters
None of these preprints announces a better battery. Together they describe something quieter: the work around the battery is getting faster and more honest. A model that predicts conductivity from 40 labeled samples cuts the cost of searching formulations by 45.6% against random search; a corrected structural model stops researchers from trusting energetics computed from the wrong reference; and an automated extraction tool with sanity checks stops bad diffusion numbers at the source. m:…3376 m:…9446 m:…3375
For anyone funding or building in batteries, the near-term gains may come less from a new chemistry than from tools that let researchers test more formulations, trust the models they use, and catch bad measurements early. The chemistry still has to work in a cell, but the path to finding out whether it works is getting shorter. m:…3376
What is still unresolved
- All three papers are preprints, so none of the results has passed peer review yet.
- Whether TabPFN keeps its accuracy on electrolyte formulations outside the AMPERE dataset is not shown.
- Whether the distorted argyrodite configuration changes predicted performance in a working cell is not addressed.
- How widely GITTcalc has been tested beyond the reported cases is not spelled out in the claims.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.