Title:	AI for Molecular Engineering
DNr:	Berzelius-2026-299
Project Type:	LiU Berzelius
Principal Investigator:	Rocio Mercado <rocio.mercado@chalmers.se>
Affiliation:	Chalmers tekniska högskola
Duration:	2026-10-25 – 2027-05-01
Classification:	10210
Homepage:	https://ailab.bio/
Keywords:

Abstract

The AI Laboratory for Molecular Engineering (AIME) at Chalmers develops machine learning methods at the intersection of artificial intelligence, chemistry, and the life sciences to engineer molecular systems for therapeutic and sustainable materials applications. Our team, led by Dr. Rocío Mercado Oropeza, focuses on developing and applying deep generative models, graph neural networks, and other AI methods to accelerate molecular design for drug discovery and materials science. This is a continuation application. Our previous allocation (3,500 GPU-h/month on Berzelius Ampere, 1,000 GPU-h/month on Berzelius Hopper, 3,000 GiB and 2,000,000 files of storage) ran to the end of September 2026. We request 4,000 GPU-h/month on Berzelius Ampere, 300 GPU-h/month on Berzelius Hopper, and 5,000 GiB with 5,000,000 files of storage for a further 6 months. The request reflects how the resource was used over the last half-year. Our compute request on Ampere is close to what was granted before (but higher to reflect our utilization), and our Hopper request is substantially reduced, because the large vision models we had expected to run there were limited during the period by storage. The storage request is where our need has increased, as several team members were unable to stage their datasets onto the cluster at all on our 3,000 GiB and 2,000,000 file allocation, and the file limit in particular proved to be the binding constraint on our imaging work, so we have applied to increase this. Our research encompasses several major projects funded by the Knut and Alice Wallenberg Foundation: (1) AI method development for targeted protein degradation (PROTACs and molecular glues), (2) time-resolved Cell Painting image analysis and generative modeling for drug discovery through the WASP/DDLS TIMED NEST, and (3) AI-driven solid polymer electrolyte design through the WASP/WISE SPEED NEST. Our computational workflows require GPU resources for training large-scale deep learning models: graph neural network- and transformer-based models for molecular property prediction, diffusion and language models for molecular generation, vision transformers for microscopy image analysis, and reinforcement learning agents for molecular optimization. We also hold a NAISS Medium allocation on Arrhenius, which we consistently exceed, and Berzelius remains important to us for single-node multi-GPU training. As of October 2026 the team comprises 11 PhD students and 3 postdoctoral researchers, of whom 9 are KAW-funded, together with co-supervised PhD students and MSc students, several of whom are also KAW-funded and involved in the projects described here.