Key Findings
CrystalJev optimizes the materials discovery funnel by replacing simple energy thresholding with calibrated probabilities and conformal guarantees. The researchers found that across 65 Matbench Discovery models, traditional threshold calls hide cost assumptions. By ordering slower computations based on their value for the decision, the system recovers most of the benefit of these computations at a tenth of the cost. In a registered prospective test involving 2,100 DFT calculations on ionic substitutions, forecasts calibrated on existing data over-stated the stable fraction of new candidates by 0.9–2.2 percentage points due to population shift.
Technical Details
The system operates as a “System-1” decision layer that processes a candidate through a frozen potential once, without relaxation. It utilizes pooled node embeddings and ten physics scalars, including single-point energy, forces, and stresses. Small heads are then trained on a random slice of labeled candidates to answer typed questions (yes/no, ordinal, choice, or compound) with calibrated probabilities. This framework allows for a specific decision rule: if a hit is worth 1 and an attempt costs c, a candidate is worth attempting when the calibrated probability p ≥ c.
Background & Context
Computational materials discovery typically follows a funnel: generative models propose millions of candidates, foundation models relax them, and stable candidates are sent to density-functional theory (DFT) or laboratory testing. Current practices often treat every candidate alike, despite the fact that a relaxation costs approximately thirty forward passes of a potential and DFT costs orders of magnitude more. CrystalJev mimics human cognitive modes, where a fast System 1 settles simple questions and a slow System 2 is engaged only when stakes justify the effort.
Significance & Outlook
By separating the state of a candidate from the question asked and the effort spent, CrystalJev provides a contract on what model outputs mean. The study indicates that while calibration can degrade under distribution shift, adding a few hundred labels from a new pool removes bias more reliably than label-free corrections.
Source: https://arxiv.org/html/2610.06985v2
Get our weekly technology intelligence — free
Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.
Subscribe Free — Weekly Tech Intelligence
By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.
- Your email and selected fields are used only to deliver the newsletter.
- We never share your information with third parties.
- You can unsubscribe anytime via the link in each email.
See our Privacy Policy for details.
Takes about a minute · Unsubscribe anytime
