Research Under Review · Nature Portfolio (npj)

CliffordIP: a new geometric language for simulating catalysts

July 27, 2026 · Submitted to a Nature Portfolio (npj) journal

Can Polat, Erchin Serpedin, Mustafa Kurban, Hasan Kurban — Texas A&M University, Ankara University & Hamad Bin Khalifa University

Read the preprint (PDF) ↗ GitHub ↗

Decarbonizing the global economy runs through catalysis: converting CO₂ back into fuels, fixing nitrogen without fossil energy, coupling carbon atoms into useful chemicals. Finding better catalysts computationally means evaluating adsorbate–surface energetics across thousands of chemically diverse surfaces — and the density functional theory calculations required are brutally expensive. Machine-learning interatomic potentials (MLIPs) promise a shortcut, but the leading architectures pay for their accuracy with heavy mathematical machinery. CliffordIP takes a different road.

Multivectors instead of spherical harmonics

Most state-of-the-art equivariant MLIPs are built on spherical harmonics and Clebsch–Gordan tensor products — mathematically elegant, computationally costly, and painful to implement. CliffordIP is a message-passing interatomic potential built instead on the Clifford algebra Cl(3,0): each atomic environment is represented as an 8-dimensional multivector — scalars, vectors, bivectors, and pseudoscalars — coupled through the geometric product. This representation natively captures bond directions, reaction planes, and local chirality, the exact geometric features that matter for catalytic intermediates, while enforcing full O(3) equivariance (rotations and reflections) via the Pin(3) group.

Results on catalysis benchmarks

On catalysis-focused adsorbate subsets, CliffordIP reduces energy mean absolute error by 20–24% relative to the next-best baseline on the three reactions that matter most for the energy transition: CO₂ reduction, N₂ reduction, and C₂ (C–C coupling) formation. Under a unified, compute-constrained protocol it also achieves the strongest in-distribution energy performance among compared baselines on the Open Catalyst OC20 and OC22 benchmarks, while remaining competitive on force magnitudes.

Why it matters

The result establishes Clifford algebra as a practical foundation for energy-focused MLIPs: less implementation overhead than spherical-harmonics stacks, with better energy accuracy where it counts. For computational catalysis groups, that translates directly into more candidate surfaces screened per GPU-hour — and a faster path from simulation to synthesizable catalyst.

Applications & economic impact

The reactions CliffordIP targets sit at the center of billion-dollar decarbonization markets: CO₂ reduction feeds e-fuels and carbon utilization, N₂ reduction underpins green ammonia and fertilizer, and C–C coupling builds the platform chemicals of a post-fossil industry. The bottleneck in all of them is screening cost:

From CPU-months to milliseconds — a single DFT evaluation of an adsorbate–surface system can consume thousands of CPU-hours; a trained MLIP answers in milliseconds, multiplying the candidate surfaces screened per GPU-hour by orders of magnitude.
Accuracy where budgets are decided — the 20–24% energy-error reduction lands exactly on the quantity (adsorption energetics) that ranks candidate catalysts, so cheaper screening does not mean worse decisions.
Drop-in adoption — the model ships as an installable package, so existing computational-chemistry pipelines can evaluate it without re-engineering.

Go deeper

Full preprint: Download the PDF — complete architecture, training protocol, and benchmark tables.

Code, checkpoints & logs: github.com/KurbanIntelligenceLab/CliffordIP

Install from PyPI: pip install cliffordip

The paper is currently under peer review; the journal version will be linked upon publication.