Microsoft Research has released Skala 1.1, an updated version of its deep-learning density functional theory (DFT) model that improves accuracy and expands integration with widely used scientific software packages. The new release, trained on 2.5 times more data than its predecessor, targets the persistent problem of balancing predictive accuracy against computational cost in chemistry simulations - a bottleneck in fields like drug discovery, materials science, and energy research.
Skala 1.1 achieves a weighted average error of just 2.8 kcal/mol on the GMTKN55 benchmark, a widely respected suite for evaluating DFT functionals. That performance rivals expensive global hybrid functionals while carrying the computational cost of a simpler meta-GGA functional, according to the Microsoft Research announcement.
This matters because DFT has long forced researchers to choose between accuracy and speed. Traditional methods often involve a trade-off, and the typical approach in the field has been to add new functionals without replacing older ones. Microsoft Research is departing from that strategy: each Skala release is designed to supersede the last, building on updated data, model architectures, and training strategies.
Integration into established scientific tools
Accuracy alone is not enough. For a computational tool to have real impact, it must fit into the workflows scientists already use - and DFT is the backbone of research across chemistry, materials science, and energy technology. Microsoft is addressing this by expanding Skala's availability into established software.
Related: AI for Science & Research
The tool is now available within CP2K, a widely used DFT simulation package, and is being integrated into other major codes such as Psi4, FHI-aims, ORCA, and VASP. Skala runs efficiently on both CPUs and GPUs with performance comparable to semi-local meta-GGAs, making it practical for large-scale simulations. This open integration strategy is critical: the success of an AI-based scientific tool depends on adoption by the developer community, not just its benchmark scores.
A living benchmark for measuring progress
Microsoft Research is also launching a "living benchmark" for Skala, an initiative that will continuously track the computational performance of successive releases across different software packages and hardware platforms. The goal is to give the community a transparent, up-to-date reference point for measuring progress.
Competition and positioning
Microsoft Research's AI for Science initiative faces competition, though direct comparisons in computational chemistry are nuanced. The company's focus on deep-learning DFT and its integration strategy distinguish it from broader AI platforms that serve different purposes.
The tool joins a growing field of specialized AI research initiatives, from other scientific simulation approaches to AI tools that support research workflows across disciplines.
Why this matters for science and research professionals
For researchers working in chemistry, materials science, or related fields, Skala 1.1 reduces the barrier to high-fidelity simulations. It no longer requires specialized hardware or deep expertise in AI model deployment to access near-hybrid-level DFT accuracy at meta-GGA cost. Smaller labs and academic groups can now tackle problems that were previously out of reach. For research teams, this translates to faster since few development cycles, more reliable predictions, and a shorter path to validating new materials or drug candidates. The iterative release strategy combined with the living benchmark means this capability, benchmark performance can be expected to improve further - and that improvement will be visible, updateable. As with any new research tool, professional development around AI techniques is another avenue to understand. AI Research Courses & Certification can help researchers apply these emerging methods in their own work.
Your membership also unlocks:
