Find the best materials for your application. AI-powered screening with honest confidence estimates.
Everything a materials scientist needs to make confident screening decisions.
Select "Solar Cell", "LED", "Thermoelectric", or "Wide-Gap Semiconductor" and the system configures target ranges automatically.
Every prediction gets an honest confidence label. Isotonic regression ensures the uncertainty bars actually mean what they claim.
Pareto sorting balances band gap, stability, and confidence simultaneously. Single-property filtering misses these tradeoffs.
Five-dimensional profiles let you compare candidates at a glance across band gap match, stability, confidence, strength, and formability.
PBE band gaps underestimate by ~40%. We label the DFT functional for each prediction. No magic claims, just transparent science.
Forward models evaluated on Matbench v0.1 with fixed splits. Screening validated by checking known good materials appear in results.
From target properties to ranked candidates in seconds.
Specify the properties you need. "I want a material with band gap 1.0 to 1.5 eV for solar cells, thermodynamically stable."
MatScreen filters 230,000 materials from Materials Project and JARVIS, applying stability and element constraints.
An ensemble of 5 ALIGNN models predicts properties for each candidate. Disagreement between models gives calibrated confidence intervals.
Pareto sorting ranks candidates across all objectives simultaneously. No single property dominates. The best tradeoffs rise to the top.
Each candidate shows predicted properties with confidence ratings. GREEN means trust it. RED means verify with DFT before acting.
Uncertainty < 0.08 eV. The model has seen many similar materials. Prediction likely within 0.1 eV of truth.
Uncertainty 0.08 to 0.15 eV. Useful prediction but should be verified with a DFT calculation.
Uncertainty > 0.15 eV. Unusual material. Do not rely on this prediction alone.
MatScreen is free and open source under the Apache 2.0 licence. Built on Materials Project, JARVIS, and ALIGNN.
View on GitHub