Weight-Space Programming for Materials Science

A materials discovery campaign often begins with almost no measurements of the property it is searching for, and the models that rank candidates or screen the output of a generator then have to be built from those first few labels. Following recent work on weight-space programming, in which a language model writes the weights of a small network directly as text, we show that a general-purpose model, given the task in words, the names of the inputs and a network of 16 hidden units with every weight set to zero, can provide this starting point by filling in all of the weights from what it has learned about materials. With no labels, the networks of the strongest model rank experimental alloys at Spearman co-efficient = 0.64 to 0.79, where the best of five simple descriptors reaches 0.47 to 0.62. They rank the bulk modulus of crystals added during generative discovery loops at 0.95, where the inverse volume per atom reaches 0.89, and 35 phase-equilibrium targets computed with Thermo-Calc at a mean of 0.67, close to the 0.70 of an extrapolation from calculated binary and ternary data. As the prior mean of a Gaussian process they raise the ranking at five labels from 0.35 to 0.69 on alloys and from 0.15 to 0.95 on crystals, and they lower the regret of Bayesian optimization significantly on 4 of 5 experimental tasks and on 13 to 16 of 33 computed targets, depending on how multiple tests are corrected, without raising it significantly on any. The same networks also serve as the mean of a Bayesian neural network and as a screen for the samples of two crystal generators, where they do as well as standard models trained on 400 labeled structures. The gain is confined to the first tens of labels and to targets that no rule or sub-system extrapolation already ranks.

Authors

Publication Details

Journal
arXiv (Cornell University)
Published
2026-11-17
DOI
https://doi.org/10.48550/arxiv.2611.22013
Primary Topic
Machine Learning in Materials Science
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Weight-Space Programming for Materials Science

Sk Md Ahnaf Akif Alvi, Raymundo Arróyave
arXiv (Cornell University)
Machine Learning in Materials Science
preprint

Weight-Space Programming for Materials Science

Sk Md Ahnaf Akif Alvi, Raymundo Arróyave
preprint en

Abstract

A materials discovery campaign often begins with almost no measurements of the property it is searching for, and the models that rank candidates or screen the output of a generator then have to be built from those first few labels. Following recent work on weight-space programming, in which a language model writes the weights of a small network directly as text, we show that a general-purpose model, given the task in words, the names of the inputs and a network of 16 hidden units with every weight set to zero, can provide this starting point by filling in all of the weights from what it has learned about materials. With no labels, the networks of the strongest model rank experimental alloys at Spearman co-efficient = 0.64 to 0.79, where the best of five simple descriptors reaches 0.47 to 0.62. They rank the bulk modulus of crystals added during generative discovery loops at 0.95, where the inverse volume per atom reaches 0.89, and 35 phase-equilibrium targets computed with Thermo-Calc at a mean of 0.67, close to the 0.70 of an extrapolation from calculated binary and ternary data. As the prior mean of a Gaussian process they raise the ranking at five labels from 0.35 to 0.69 on alloys and from 0.15 to 0.95 on crystals, and they lower the regret of Bayesian optimization significantly on 4 of 5 experimental tasks and on 13 to 16 of 33 computed targets, depending on how multiple tests are corrected, without raising it significantly on any. The same networks also serve as the mean of a Bayesian neural network and as a screen for the samples of two crystal generators, where they do as well as standard models trained on 400 labeled structures. The gain is confined to the first tens of labels and to targets that no rule or sub-system extrapolation already ranks.

arXiv (Cornell University)
Machine Learning in Materials Science
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon