Perspective: Towards sustainable exploration of chemical spaces with machine learning

A major collaboration of 23 scientists from leading institutions has published a critical perspective on the unsustainable computational practices in AI-driven materials and drug discovery. The paper, stemming from the 'SusML' workshop in Dresden, argues that the field's reliance on generating enormous quantum-mechanical (QM) datasets for training models incurs prohibitive energy and infrastructure costs. This 'brute-force' approach, while enabling progress, creates a significant environmental and economic bottleneck for scaling discovery.

The authors propose a strategic shift towards hierarchical, multi-fidelity workflows. This involves using fast, general-purpose machine learning models (ML surrogates) for broad screening and reserving expensive, high-accuracy QM calculations only for the most promising candidates. They highlight key efficiency strategies like active learning, where the AI decides which data points are most valuable to compute next, and model distillation, which compresses large models into smaller, more efficient versions.

Crucially, the perspective emphasizes that sustainability must be coupled with real-world practicality. This means AI models must account for synthesizability and multi-objective design criteria—not just theoretical performance—to ensure discoveries can be physically realized. The researchers advocate for open data, reusable workflows, and domain-specific AI systems as foundational pillars for a new era of efficient and responsible scientific discovery.