Atomistic simulation requires three issues: data of the science, compute-efficient implementation of simulations, and accessible interfaces to the simulation stack.
The primary stays the researcher’s area, as no software substitutes for understanding what to simulate or recognizing a bodily significant outcome. NVIDIA ALCHEMI Toolkit, launched earlier this yr, has dramatically lowered the second barrier for Machine Studying Interatomic Potentials (MLIP) with composable, PyTorch-native constructing blocks for establishing GPU-accelerated simulation workflows with in-flight batching enabled.
The third barrier has continued. Not like classical drive fields, the MLIP ecosystem remains to be nascent, and the accessible interfaces that exist for classical simulations are very restricted. They run on a distinct software program stack than the instruments many computational chemists are accustomed to with new information constructions, composition patterns, and dependencies.
AI coding brokers supply a method by: They generate and execute code from natural-language descriptions written within the phrases a researcher would use in every day technical discussions. However a general-purpose agent could not know the ALCHEMI Toolkit API, and might produce plausible-looking code that solely seems to make use of it appropriately.
ALCHEMI Toolkit agent abilities and reference information present the lacking API patterns on demand, leaving the immediate to your science: the fabric, the situations, and the constraints on the simulation protocol.
Constructing simulation workflows with a coding agent
This publish follows an end-to-end ALCHEMI Toolkit workflow: what the researcher begins with, how they immediate the agent, what code and simulation pipeline it produces, and the way the outcomes are validated on NVIDIA H200 GPUs. It additionally distills classes from 45 generated pipelines into sensible steerage for constructing reliable GPU-accelerated simulation workflows with coding brokers.
Tips on how to get began
System and bundle necessities
Python ≥3.11, <3.14
PyTorch ≥2.8
CUDA 12 or CUDA 13, with a appropriate NVIDIA driver (570+ advisable)
Working System: Linux (major), macOS
NVIDIA GPU (RTX 20xx or newer), CUDA Compute Functionality ≥ 7.0
Minimal 4 GB RAM (16GB advisable for big methods)
Set up
The way you arrange the agent’s surroundings meaningfully impacts the reliability of generated code. We advocate putting in the Toolkit in a runnable Python surroundings and letting the agent execute the scripts it generates. Within the remaining 45-pipeline marketing campaign, this setup produced no damaged imports or references to nonexistent APIs.
Step 1: Create a Python surroundings and set up ALCHEMI Toolkit with uv bundle supervisor:
# Create native surroundings at .venv
uv venv –seed –python 3.12
# Set up ALCHEMI Toolkit into .venv
uv pip set up “nvalchemi-toolkit[mace,ase]==0.2.0”
To run on an NVIDIA GPU, embrace the CUDA additional that matches your CUDA surroundings. For instance, for CUDA 13, set up nvalchemi-toolkit[mace,ase,cu13]==0.2.0.
Step 2: Obtain agent abilities from the identical launch tag in order that they match the put in API:
# Obtain the nvalchemi-toolkit abilities
npx degit NVIDIA/nvalchemi-toolkit/.claude/abilities#v0.2.0 .claude/abilities
Step 3: Set up a coding agent. This benchmark used Claude Code:
npm set up -g @anthropic-ai/claude-code
# Or natively by curl
curl -fsSL https://claude.ai/set up.sh | bash
# Begin a claude session within the native folder
claude
Open the agent in your mission listing, permit it to execute code, and begin describing simulations; the agent masses the related abilities on demand. Letting the agent run what it writes removes virtually all mechanical errors earlier than you see the script. From there, the standard is determined principally by your immediate.
With out a runnable shell, a supply git-checkout is a fallback: In earlier assessments, studying the supply eradicated damaged imports throughout 617 import statements. A pip set up with neither shell nor supply was the weakest configuration.
Any agent supporting the open Agent Abilities commonplace works, together with Cursor and OpenCode (see the Agent Abilities consumer information for configuration).
Greatest practices for writing a very good immediate
The next rules are synthesized from measurable variations within the benchmark’s output high quality, not from basic coding-agent recommendation. The 5 immediate ranges, from least to most specified, are proven in Determine 1, beneath.
Identify the system, methodology, and scale all the time; add a CLI contract just for unattended operation. Prompts naming the fabric, methodology, and scale scored greatest. Full CLI contracts achieved full reusability, however value about 4× extra tokens and produced 2.3× extra code than Sketch prompts. Spec prompts have been most fragile, accounting for 3 of seven screening failures. Specify the science and deliverables; let the talents provide API patterns.
Identify the fabric, part, and reference conference explicitly. Underneath-specifying the system induced the clearest physics failures in earlier assessments: “a transport property of a Li materials” produced an argon demo, and two Cu scripts used completely different adsorption references, materially affecting the outcomes. The ultimate prompts get rid of these failures by naming the fabric, part, and reference conference. Protocol issues too: Scripts and not using a thermostat instruction used Langevin manufacturing dynamics that damped diffusion by 3-5×; requesting NVE modified each script to the suitable measurement ensemble.
Specify the constraint, not the implementation. Describe what the script should accomplish relatively than inside API courses. In a managed comparability, naming a pipeline assemble modified none of 12 implementations. The API sample comes from the examples and abilities, not the immediate. Naming internals and not using a full interface contract is the place fragility concentrates.
Ask explicitly for self-assessment and premise-checking. Brokers didn’t query whether or not any requested property was bodily well-posed. Ask explicitly for premise checks, validation, and uncertainty estimates, and require restoration of an impartial identified outcome; the agent received’t add these by itself.
Three workflows, from immediate to GPU execution
The next workflows illustrate what ALCHEMI Toolkit and a coding agent can deal with end-to-end from a natural-language immediate. To floor these suggestions in measured outcomes, we ran a scientific benchmark of 45 simulation pipelines throughout three workflows, 5 immediate ranges, and three samples per degree.
Silicon equation of state (EOS)
Oxygen adsorption on Cu(111)
Li self-diffusion molecular dynamics (MD)
Scripts have been assessed two methods. Firstly, deterministic code-feature checks: property protection (proper amount, proper method), API-pattern protection (use of the toolkit’s batched API floor), and reusability (a parameterized, re-runnable interface). Secondly, execution on an identical NVIDIA H200 GPUs as floor reality.
All 45 scripts used batched GPU execution. L2-L5 prompts required batching; L1 adopted it from the Toolkit abilities and examples. We screened all scripts at demonstration scale and ran one consultant per workflow and degree at manufacturing settings.
Bulk silicon equation of state
The duty: The equation of state of diamond-cubic silicon
The energy-volume curve whose minimal offers the lattice fixed a0 and whose curvature offers the majority modulus B0. At each immediate degree the agent constructed the identical pipeline: 50-60 strained volumes relaxed concurrently as one GPU batch, then a Birch-Murnaghan match. All 5 manufacturing representatives conform to the final digit, a0 = 5.4661 Å and B0 = 88.15 GPa, contained in the established all-electron PBE reference window. The lattice fixed differs from the NIST experimental worth, whereas the bulk-modulus comparability comes from McSkimin’s experimental measurements. These offsets are in line with reported PBE conduct. The takeaway mirrors the benchmark-wide sample: immediate specificity modified the code’s construction and price, by no means the physics.
Oxygen adsorption on Cu(111)
The duty: Rank the adsorption websites for atomic oxygen on Cu(111)
The pipelines freeze the slab’s backside layers, chill out 24+ candidates throughout the 4 high-symmetry websites as one GPU batch, and compute Eads = Eslab+advertisements − Eclean slab − EO. Each manufacturing consultant finds the fcc hole most steady, Eads(fcc) = −4.799 ± 0.004 eV throughout the 5 ranges, with the ordering fcc ≥ bridge > hcp >> prime. Reported reference values are −4.31 eV at 0.25 monolayer. Copper microcalorimetry research by Giamello et al. and Naumann d’Alnoncourt et al. reported values from −4.46 to −4.60 eV. Protection variations and mannequin error could contribute to the stronger binding. Millielectronvolt settlement throughout impartial scripts is what a said reference conference buys.
Lithium self-diffusion through molecular dynamics
The duty: Estimate the self-diffusion coefficient D of liquid lithium.
The pipelines construct a bcc lithium supercell, soften and equilibrate it above lithium’s 454 Ok melting level, propagate three temperatures occasions three seeds (9 replicas) as a single batched system on the GPU, and extract D from the mean-squared displacement, MSD — the common squared distance that atoms journey — whose slope over time offers D by the Einstein relation.
Technical specs
The next configuration was used to provide the workflow examples on this publish:
Coding agent: Claude (claude-opus-4-8)
Effort degree: Excessive
MLIP checkpoint: MACE-MPA-0 (medium-mpa-0)
Toolkit model pin (0.2.0)
Benchmark insights
The benchmark produced three sensible findings by systematically various immediate degree, software entry, and execution functionality.
Immediate specificity buys code construction. Property protection was 1.00 at each degree, so the science is correct from the primary immediate. What specificity introduced was construction: API-pattern protection doubles at L4-Spec (0.52 to 0.96) and full interface reusability arrives solely with the L5-Contract (0.67 to 1.00) (Determine 7, beneath).
Brokers defaulted to acquainted algorithms: 38 of 45 scripts used FIRE and none used FIRE2, regardless of FIRE2 being documented beside it as an improved variant. What the examples and abilities showcase, the agent makes use of; what they don’t, it fills from pretraining.
Execution uncovered failures that CPU self-tests couldn’t. The runnable shell eradicated damaged imports, however seven scripts nonetheless failed on GPU-specific paths. Extra elaborate prompts exercised extra of those paths and failed extra usually. Self-testing improves code technology, however doesn’t change validation on course {hardware}. A technology sandbox with GPU entry could shut this hole, however our methodology accounts for what a typical consumer would do in apply: Use “Claude autorun” within the Claude app on their laptop computer after which undergo a GPU cluster after.
Limitations
Scientific judgment stays important. Throughout each examined situation, zero configurations pushed again on a bodily ill-posed job. In an earlier probe, each agent complied with a request for Li-ion diffusion in pristine crystalline LiF, a property not measurable on the simulated timescale. Nonetheless, it’s doable a web-equipped agent would run this test; the technology sandbox had no net entry, so the agent couldn’t have regarded up whether or not the property was believable. Brokers additionally selected believable timesteps and thermostats with out assessing their suitability for the fabric or timescale. Brokers speed up translation from scientific intent to code, however can not decide whether or not the meant simulation is bodily significant.
MLIP basis fashions are usually not universally correct. The benchmark used MACE-MPA-0 (medium-mpa-0), whose accuracy varies exterior its coaching distribution. Our silicon, copper, and liquid-lithium outcomes are cheap towards their references, however each new chemistry must be validated towards DFT or experiment. It is a mannequin and generally reference DFT limitation, not an ALCHEMI Toolkit or agent limitation.
At all times evaluate numerical outcomes with an impartial reference. The Langevin-production lithium scripts reported diffusion coefficients that have been statistically impeccable but damped 3-5× by the thermostat; solely comparability towards the temperature-matched experimental anchor exposes it. The test should even be impartial: one script validated its unit conversion by producing artificial information with the identical fallacious fixed, so its self-test handed. Statistically dependable and bodily significant could be orthogonal.
Ecosystem adoption
A number of firms have been exploring the approaches we focus on on this weblog. In Might 2026, Matlantis, an AI-powered simulation platform utilized by over 150 firms and an ALCHEMI ecosystem accomplice, launched a public Abilities library on GitHub and introduced a Claude Code integration for its common simulator. BIOVIA can also be leveraging NVIDIA ALCHEMI NIM in Dassault Systèmes’s AI-powered digital companion, MARIE, to democratize entry to superior simulation capabilities and allow handy execution of molecular simulations, empowering extra scientists to develop and validate Digital Twins of supplies.
Go to the NVIDIA/nvalchemi-toolkit GitHub repo and NVIDIA ALCHEMI Toolkit documentation.
The repository contains agent abilities and greater than 30 workflows throughout primary, intermediate, superior, and distributed use instances.
AcknowledgmentsWe wish to thank Nikita Fedik, Susumu Ohno from Matlantis Company, and James Wescott from Dassault Systèmes – BIOVIA for his or her contributions to this publish.
