Same 2.4 eV band gap, different routes: SHAP explains per-prediction feature contributions
bravo_abad · x · 2026-11-01
An explainability explainer for materials-science ML: a model can predict the same 2.4 eV band gap for two crystals for very different reasons.
SHAP (SHapley Additive exPlanations) starts from a reference prediction (usually the model's average over a background dataset) and assigns each descriptor an eV contribution that sums exactly to the final prediction. In the example with a 1.6 eV reference, material A is pushed up by electronegativity (+0.9) and d-band width (+0.5), while material B is pushed up by coordination (+0.8) and cell volume (+0.6) — same 2.4 eV, different decomposition.
Global feature-importance rankings can't reveal this per-prediction difference; the accompanying chart visualizes both waterfall decompositions.
More from Research
- CrowdGPT v2.6: crowdsourced LLM training adds 4.5B tokens in 6 days — Vxtzq1 · 2026-11-01
- AI agent performs first end-to-end epilepsy treatment in a mouse, from gene therapy design to robotic brain surgery — Distinct-Question-16 · 2026-11-01
- Interactive Explainer Demystifies Gated DeltaNet, the Fixed-State Linear Attention Behind Recent Models — CatAstro_Piyush · 2026-11-01
- BooBench: A Halloween Benchmark That Scares LLMs and OCR Tools Into Character-Repeating Meltdowns — xennygrimmato_ · 2026-11-01
- Interactive visualizer lets you zoom into Esperanto's ET-SoC-1 AI chip, down to FMA instructions — yaroslavvb · 2026-11-01
- Sam Altman admits OpenAI let its AI-generated math research get 'out of control for one week' — Polymarket · 2026-11-01