A Physics-Constrained AI Agent Speeds Battery Interface Analysis

Direct answer

What does the battery XPS agent automate?

It uses a local large-language-model planner to route XPS data through deterministic readers, physically constrained peak fitting, quantification, plotting and hyperspectral mapping. The paper reports more than 140-fold faster processing in its demonstrations, but the framework does not by itself prove that every chemical assignment or new material conclusion is correct.

What happened

Zhenzhen Yang, Gabriel M. Veith, Seoung-Bum Son and Brian J. Ingram have published an open-source agentic framework for automated X-ray photoelectron spectroscopy analysis in Energy and AI. The target is a persistent bottleneck in battery research: characterising the chemically heterogeneous solid-electrolyte interphase.

Manual XPS peak fitting can be slow and analyst-dependent. The authors report a workflow that processes spectra more than 140 times faster in their demonstrations, applies repeatable templates, and combines peak fitting with principal-component analysis, k-means clustering and multivariate curve resolution for hyperspectral maps.

How AI was actually used

The language model is a planner, not the numerical engine. It receives a request, classifies the data path and calls specialised Python tools for reading, fitting, quantification, plotting and mapping. Physical rules—including binding-energy ranges, peak widths and spin-orbit relationships—are encoded in human-readable templates.

A mandatory first phase checks file integrity, energy axes, formats and data quality. Low-confidence or suspicious data can trigger a human review checkpoint. The second phase runs a constrained tool sequence. The authors evaluated routing prompts across six open-source models from 3B to 8B parameters and report mean tool-calling accuracy from 80% to 97%, with clarification the hardest category.

Why it matters

Battery research increasingly produces large spectroscopy datasets that are expensive to inspect manually. Faster, standardised analysis could make high-throughput screening more useful and could expose minority phases in spatial maps that simple integration misses.

The architecture is also instructive beyond XPS. It keeps flexible natural-language orchestration separate from deterministic scientific computation. That makes the action space, templates, logs and numerical outputs available for audit rather than asking an LLM to invent the analysis itself.

What remains unproven

Speed and reproducibility do not guarantee chemical truth. Peak deconvolution can remain non-unique, reference energies and sensitivity factors vary by instrument, and template constraints can encode an analyst's assumptions as efficiently as they encode expertise. The paper notes that users must supply instrument-appropriate relative sensitivity factors.

The demonstrations do not establish autonomous materials discovery from acquisition through validated performance. Multi-modal links to cycling efficiency, capacity fade and electrolyte formulation are presented as future capabilities. The most difficult prompt category—choosing to clarify rather than act—also showed wide model variation.

Independent verification should recreate the exact data and software environment, compare fits with expert and vendor baselines, test alternative templates and starting conditions, inspect residuals, and confirm that the claimed speedup uses comparable work. Haivian has not independently audited the framework.

Primary sources and version boundary

• Yang et al., Energy and AI 25 (2026) 100796, doi:10.1016/j.egyai.2026.100796.
• Author-posted SSRN record, posted 9 April 2026.

For automated scientific workflows that need their data, templates, numerical outputs and claims traced independently, see Haivian's AI research audit and computational reproducibility audit.