TL;DR: Skala is a deep-learning exchange–correlation functional 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 the single-reference subset) and WTMAD-2 ≈ 3.89 kcal/mol on GMTKN55; evaluations use a fixed D3(BJ) dispersion correction. This functional is currently positioned to be used for molecular chemistry in the main group, and transition metals as well as periodic systems are planned extensions. Azure AI Foundry Azure AI Foundry Labs has released the model and tools for free. microsoft/skala repository.
What is the maximum compression ratio you can achieve by using a format-aware diagram compressor? Microsoft Research releases Skala, a neural exchange–correlation (XC) functional for Kohn–Sham Density Functional Theory (DFT). Skala is able to learn non-local effects while maintaining a computational profile similar to meta-GGA functions.
Skala (and what it’s not)?
Skala uses a standard meta GGA grid to evaluate a neural form instead of a custom-made XC. It specifically Does not include This first release is an attempt to teach dispersion. Benchmark evaluations use fixed benchmarks. D3 Correction (D3(BJ), unless otherwise noted). It is not the goal to have a functional that can be used for every regime on day 1, but rather a thermochemistry of main groups at comparatively low cost.
Benchmarks
The following are some of the ways to get in touch with us. W4-17 atomization energiesSkala Reports Mae 1.06 kcal/mol The full set is available. 0.85 kcal/mol The single-reference set. The subset of single-reference GMTKN55, Skala achieves WTMAD-2 3.89 kcal/molCompared to top hybrids, all functionals have been evaluated at the same setting (D3(BJ), except when VV10/D3(0) apply).
Architecture and training
Skala aggregates data via an agregation grid after evaluating meta-GGA on the numerical integration grid. Non-local, finite-range neural operator (bounded enhancement factor; exact-constraint aware including Lieb–Oxford, size-consistency, and coordinate-scaling). Pre-training is followed by two phases of training: B3LYP densities XC Labels Extracted from High-Level Wavefunction Energies SCF-in-the-loop fine-tuning Skalas Purchase Densities are not allowed (no SCF backprop).
The model was trained by a large curated corpus that is dominated primarily by ~80k high-accuracy total atomization energies (MSR-ACC/TAE) plus additional reactions/properties, with W4-17 The following are some examples of how to get started: GMTKN55 Removing them from the training program to reduce leakage.
Estimated cost and timeframe for implementation
Skala keep Semi-local Cost Scaling It is optimized for GPU-based execution. GauXCThe public repo exposes (i) PyTorch Implementation microsoft-skala PyPI Package with PySCF/ASE Hooks (ii) and GauXC Add-on The README lists the steps to use Skala in other DFT stacks. README list There are 276 parameters. There are minimal examples.
Approval
In practice, Skala slots into The main group molecular Workflows that require semi-local accuracy and high-throughput reaction energetics (ΔE, barrier estimates), Most radical/conformist stability Ranking is a ranking system. geometry/dipole predictions feeding QSAR/lead-optimization loops. It’s revealed via PySCF/ASE The a GauXC The GPU Path allows teams to run batch SCF and screen candidates near Meta-GGA, then reserve hybrids/CC to do final checks. Skala, a tool for managing experiments and sharing data is now available. Azure AI Foundry Labs Open GitHub/PyPI Stack.
The Key Takeaways
- Performance: Skala achieves Mae 1.06 kcal/mol On W4-17 (0.855 on the subset of single reference) WTMAD-2 3.89 kcal/mol On GMTKN55, dispersion will be applied via D3(BJ) In reported evaluations.
- Method: The neural XC with meta-GGA inputs. Learned finite range non-locality, honoring key exact constraints; retains semi-local O(N³) The cost is not included in the dispersion of this release.
- Train the Signal: 150k High-accuracy labels including80k CCSD (T)/CBS quality atomization energies; SCF-in-the-loop Skala densities are used for fine tuning; the public test set is de-duplicated.
Skala: A functional neural XC reporting tool Mae 1.06 kcal/mol On W4-17 (0.855 on single reference) WTMAD-2 3.89 kcal/mol GMTKN55 evaluated using D3(BJ) Dispersion today, and the scope of The main group molecular systems. Test it out via Azure AI Foundry Labs With code and integrations PySCF/ASE on GitHub, enabling head-to-head benchmarks with existing meta-GGAs or hybrids.
Take a look at the Technical Paper, GitHub Page The following are some examples of how to get started: technical blog. Check out our GitHub Page for Tutorials, Codes and Notebooks. Also, feel free to follow us on Twitter Join our Facebook group! 100k+ ML SubReddit Subscribe now our Newsletter. Wait! What? now you can join us on telegram as well.
Asif Razzaq, CEO of Marktechpost Media Inc. is a visionary engineer and entrepreneur who is dedicated to harnessing Artificial Intelligence’s potential for the social good. Marktechpost was his most recent venture. This platform, which specializes in covering machine learning and deep-learning news, is both technically solid and understandable to a broad audience. This platform has over 2,000,000 monthly views which shows its popularity.
