Microsoft Research releases Skala: a deep learning exchange correlation function that achieves hybrid-level accuracy at semi-local cost

Long story short: Skala is a deep learning exchange correlation function for Kohn-Sham density functional theory (DFT) that targets hybrid-level accuracy at semi-local cost, reporting MAE ≈ 1.06 kcal/mol on W4-17 (0.85 on a single reference subset) and WTMAD-2 ≈ 3.89 kcal/mol on GMTKN55; evaluated using fixed D3(BJ) dispersion correction. It positions main group molecular chemistry today, with transition metals and periodic systems as future extensions. Azure AI Foundry The model and tools are now available through the Azure AI Foundry lab and open source platform microsoft/skala repository.

How much compression ratio and throughput can you recover by training a format-aware graph compressor and passing only self-describing graphs to a universal decoder? Released by Microsoft Research scalethe neural exchange correlation (XC) functional of Kohn-Sham density functional theory (DFT). Skala learns non-local effects from data while keeping computational profiles comparable to meta-GGA functionals.

What Skala is (and isn’t)?

Skala replaces handcrafted XC forms with neural functions evaluated against standard meta-GGA grid features. it explicitly No Attempts to learn dispersion in first version; baseline evaluation uses fixed D3 Correction (D3(BJ) unless otherwise stated). The goal is to achieve rigorous main-group thermochemistry at semi-local cost, rather than a universal functional for all systems on day one.

Benchmark

exist W4-17 Atomic energyscale report MAE 1.06 kcal/mol in full set and 0.85 kcal/mole On a single reference subset. exist GMTKN55Scala reached WTMAD-2 3.89 kcal/molcompeting with top hybrids; all functionals are evaluated using the same dispersion setting (D3(BJ), unless VV10/D3(0) applies).

Structure and training

Skala evaluates meta-GGA features on a standard numerical integration grid and then Limited range, non-local neural operators (Bounded enhancement factor; precise constraint perception, including Lieb-Oxford, size consistency and coordinate scaling). The training is conducted in two stages: (1) Pre-training B3LYP Density Has XC tags extracted from advanced wave function energy; (2) SCF in-the-loop fine-tuning using a balance own Density (without backpropagation through SCF).

The model is trained on a large, carefully curated corpus of ~80k high-precision total atomization energy (MSR-ACC/TAE) plus additional reactions/features, W4-17 and GMTKN55 Removed from training to avoid leakage.

Cost overview and implementation

scale maintained Semi-local cost scaling and designed for GPU execution High XC;The public repurchase agreement is made public: (i) a torch implementation and microsoft-skala The PyPI package contains PySCF/ASE hook, and (ii) GauXC plugin Can be used to integrate Skala into other DFT stacks. Readme file list ~276k parameters and provides minimal examples.

application

In practice, Skala inserts main race molecule Workflows where semi-local cost and hybrid-level accuracy are important: high throughput reaction energetics (ΔE, barrier estimate), Conformational/radical stability ranking, and Geometry/Dipole The predictions are fed back into the QSAR/lead optimization loop. because it is exposed through PySCF/ASE and a High XC In the GPU path, teams can run batch SCF jobs and filter candidates close to the meta-GGA run, then retain hybrid/CC for final inspection. For hosting experiments and sharing, Skala is available Azure AI Foundry Lab and as an open GitHub/PyPI stack.

Main points

• Performance: Scala’s achievements MAE 1.06 kcal/mol on W4-17 (0.85 on the single reference subset) and WTMAD-2 3.89 kcal/mol On GMTKN55; dispersion is by applying D3(Beijing) in the evaluation of the report.
• method: Neural XC function with meta-GGA input Non-locality in limited-scope learningobeying key precision constraints; retain Semi-local O(N3) cost and dispersion is not learned in this version.
• Training signal: Trained by ~150,000 High-precision labels including ~80k CCSD(T)/CBS quality atomization energy (MSR-ACC/TAE); SCF in the ring Fine-tuning uses Skala’s own density; the public test set removes duplicates from training.

Skala takes a pragmatic step: NeuralXC feature report MAE 1.06 kcal/mol W4-17 (0.85 for single reference) and WTMAD-2 3.89 kcal/mol On GMTKN55, the evaluation is D3(Beijing) Scattered, today’s range is main race molecule system. This can be tested by Azure AI Foundry Lab Direct head-to-head baselines against existing meta-GGAs and hybrids via code on GitHub and PySCF/ASE integration.

Check Technical papers, GitHub pages, and technical blogs. Please feel free to check out our GitHub page for tutorials, code, and notebooks. In addition, welcome to follow us twitter And don’t forget to join our 100k+ ML SubReddit and subscribe our newsletter. wait! Are you using Telegram? Now you can also join us via telegram.

Asif Razzaq is the CEO of Marktechpost Media Inc. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of artificial intelligence for the benefit of society. His most recent endeavor is the launch of Marktechpost, an artificial intelligence media platform that stands out for its in-depth coverage of machine learning and deep learning news that is technically sound and easy to understand for a broad audience. The platform has more than 2 million monthly views, which shows that it is very popular among viewers.