What happened
Microsoft Research announced Skala 1.1, a deep-learning exchange-correlation functional for density functional theory. The company says the model was trained on 2.5 times more data than the first public version, improves results across several molecular-simulation tasks, and is available in CP2K. Integrations with Psi4, FHI-aims, ORCA and VASP are in progress, alongside a living benchmark for tracking performance across software and hardware.
Microsoft Research published the announcement on August 20, 2026, describing Skala 1.1 as the latest version of a deep-learning exchange-correlation functional intended to improve density functional theory, or DFT. DFT is used to approximate electronic structure in calculations involving chemistry, materials, catalysis, energy technologies and drug discovery. The source says Skala follows a continuous-improvement model in which each release is intended to supersede the previous one while preserving roughly the same practical computational cost. Those descriptions are claims made by Microsoft Research; the supplied source does not include independent confirmation from outside evaluators.
The company reports that Skala 1.1 was trained on 2.5 times more data than the first public Skala release. Microsoft says the additional Microsoft Research Accurate Chemistry Collection data added categories including electron affinities and noncovalent clusters, increasing both the amount and diversity of the training set. According to the source, the updated model improves performance in main-group thermochemistry, reaction kinetics and molecular-structure prediction, while also producing accurate electron densities, dipole moments and molecular geometries. The announcement does not provide a complete accounting of the training data, its licensing, or the computational resources used to create the model.
On the GMTKN55 benchmark suite, which the source describes as covering 55 categories of chemical problems, Microsoft reports a weighted average error of 2.8 kilocalories per mole. It says Skala 1.1 ranked first in 32 categories and outperformed leading global, including range-separated, hybrid functionals while operating at the cost of a semi-local meta-GGA functional. In a separate implementation check, Microsoft says CP2K and PySCF results agreed within 0.1 kilocalories per mole mean absolute deviation across a representative GMTKN55 subset, with one outlier involving a difficult radical system. The supplied material does not independently verify these figures or show the full comparison methodology.
The access announcement has two parts. Skala was previously released through an open-source community edition built on GPU4 PySCF and integrated with ASE, according to Microsoft. The company now says Skala is available in the open-source CP2K package following collaboration with the Center for Advanced Systems Understanding, while work continues on Psi4. Microsoft also says it is working with developers of FHI-aims, ORCA and VASP, but describes those efforts as integrations in progress rather than completed availability. A benchmarking harness and living performance report are intended to track successive releases across packages and hardware platforms.
Read the primary source: microsoft.com ↗
Why it matters
Density functional theory is widely used to estimate molecular and material properties, but accuracy and computational cost have traditionally been difficult to balance. If Microsoft’s reported benchmark results transfer across implementations and practical workloads, Skala could give researchers access to higher-accuracy calculations without the full cost of some hybrid functionals. The announcement matters chiefly because it combines a model update with broader software access and reproducible performance tracking.
The technical significance is the attempt to narrow a long-standing tradeoff in electronic-structure calculations. More accurate methods can be costly, limiting the size or duration of simulations. Microsoft’s central claim is that Skala 1.1 reaches accuracy associated with more expensive global hybrid functionals while retaining the computational profile of a semi-local meta-GGA. If independently reproduced, that combination could allow some researchers to run more accurate calculations at a scale that would otherwise require compromises in system size, simulation time or hardware.
Access through established chemistry packages could matter as much as the model itself. Researchers often build workflows around particular electronic-structure codes because those packages support different methods, system sizes and forms of molecular or materials simulation. Microsoft says CP2K is especially suited to large-scale systems and long-timescale molecular dynamics, and that Skala is being brought to other widely used platforms. Broader native integration could reduce the need for researchers to rebuild workflows around a separate implementation, although the announcement does not quantify installation effort, user uptake or the degree of feature parity among packages.
The reported results could be relevant to scientific areas where approximate molecular energies and structures guide decisions about which experiments or higher-accuracy calculations to pursue. Faster or more accurate calculations might help screen candidate molecules, materials or reaction pathways. That is a potential research benefit, not an outcome established by this announcement: Microsoft provides benchmark and implementation results but does not document a new drug, material, industrial process or experimentally validated discovery produced with Skala 1.1.
The living benchmark also addresses a practical problem in evaluating scientific software. Performance can change when a model, numerical library, compiler, hardware platform or implementation changes. A continuously updated report could make those changes easier to compare than a single published timing result. Its value will depend on whether the harness is sufficiently transparent and whether comparisons use representative workloads, consistent accuracy requirements and clearly reported hardware. The source announces the resource but does not provide enough information here to assess its coverage or governance.
What to watch next
The main unanswered questions are whether independent researchers reproduce the reported accuracy, how Skala performs outside GMTKN55 and the cited test subset, and when the planned integrations become broadly usable. The living benchmark may provide useful evidence if it includes transparent hardware, software-version and workload details over time. The source does not establish real-world discoveries, adoption levels, licensing terms across all packages, or performance for every system size and chemistry.
Independent replication should be the first test. The announcement is from Microsoft Research, the organization developing Skala, and all accuracy and performance claims in the supplied material come from that source.
Researchers will need to confirm the GMTKN55 results, inspect the treatment of difficult cases such as radicals, and determine whether the reported 2.8-kilocalorie-per-mole weighted average error is comparable under the same settings used for competing functionals. Results on additional benchmark suites would help show whether the gains generalize beyond the reported evaluation.
Implementation consistency is another key issue. Microsoft reports that CP2K and PySCF agreed within 0.1 kilocalories per mole mean absolute deviation across a representative subset, but that is not the same as demonstrating identical behavior across every supported system, basis set, convergence setting or hardware configuration. The planned Psi4 integration and the work involving FHI-aims, ORCA and VASP should be watched for completion, documentation, version compatibility and reproducible tests. Availability may differ by package, and the source does not specify a common release date for all integrations.
The performance claims need to be interpreted by workload. Microsoft says Skala has performance comparable to semi-local meta-GGAs on CPUs and GPUs, that GPU cost matches r2SCAN in the cited comparison, and that CPU overhead disappears for molecules larger than roughly 20 to 30 atoms or systems larger than roughly 300 orbitals. These are source-reported comparisons, not universal guarantees. Researchers will need timings for their own molecular sizes, hardware, software stacks and simulation types, especially because small systems may carry overhead and different packages may scale differently. Finally, the living performance report should reveal whether later Skala releases improve accuracy without creating hidden cost or instability. The source says new releases and optimizations in libraries such as GauXC can change efficiency, and that the benchmark will track this progress. Important unknowns include the model’s licensing and support arrangements across each package, the long-term maintenance of integrations, the range of chemical environments represented in training data, and whether the method remains reliable when used for decisions outside the benchmark conditions. No adoption, industrial impact or experimentally confirmed discovery is established yet.
