New paper on high-entropy alloy inverse design

How can we design high-entropy alloys for a target property without testing every possible composition?

Our new paper is now available on arXiv:
QUBO-Compatible Active Learning for Inverse Design of High-Entropy Alloys
Read the full paper: https://arxiv.org/abs/2608.28239

High-entropy alloys offer a vast number of possible element combinations and mixing ratios, making it difficult to identify compositions that maximise a desired property. Our work approaches this hurdle as an inverse-design problem, where we map the training data (alloy composition and bulk modulus label) to a binary latent space - you can think of this as an n-dimensional relief map, where the height of the surface corresponds to the bulk modulus. We then navigate this map to find the highest point (i.e. the largest bulk modulus) using active learning, and from there find the corresponding alloy composition.

In more technical language: The workflow combines a pretrained graph neural network, a binary variational autoencoder, and active learning. The autoencoder maps alloy compositions into binary latent codes, and a factorization machine then learns which codes are associated with high predicted bulk moduli.

A unique feature of our workflow is that the resulting surrogate model is quadratic in binary variables and can be exported directly as a QUBO problem. Our benchmark shows that our workflow remains competitive with strong classical optimization methods; in particular, local changes around already promising candidates were very effective for discovering high-scoring alloy compositions.

The key result is a workflow that combines targeted, data-driven alloy design with a direct QUBO-compatible optimization endpoint. This work was developed by HQS Quantum Simulations together with the Institute for Frontier Materials on Earth and in Space at DLR.