For centuries, discovering a new material took decades of lab work. Alchemist is our gateway to NVIDIA ALCHEMI: atomistic simulation with near-quantum-chemistry accuracy at GPU speed, running on-premise on a DGX Spark. Thousands of molecular candidates evaluated before touching a single flask.
* Figures published by NVIDIA (SC24 / SC25)
Computational chemistry has always lived a war between accuracy and speed: DFT is exact but painfully slow; classical force fields are fast but imprecise. ALCHEMI breaks that trade-off with GPU-accelerated machine-learning interatomic potentials (MLIPs), across four integrated pillars:
Exploring the materials universe and proposing new candidates from target properties.
Interatomic potentials with near-quantum-chemistry (DFT) accuracy at a fraction of the computational cost.
Pretrained foundation models mapping molecular representations to properties for fast screening.
Simulation tooling to generate training datasets and fine-tune new use cases.
Batched Geometry Relaxation — batched relaxation (MACE-MP-0 / AIMNet2)
Batched Conformer Search — low-energy conformers
Batched Molecular Dynamics — high-throughput MD
Batched quantum validation
Warp kernels: MD, FIRE, Ewald/PME, >100,000 atoms
SC24: announcement + BGR · SC25: BCS, BMD and Batched DFT
Each case is a screening that today takes months of lab work or supercomputing clusters — and in this PoC runs in hours, on a machine that fits on your desk.
Batched screening of thousands of electrolyte formulations, SEI additives and selective sorbents to extract lithium from brine without evaporation ponds. The natural case for Chile, heart of the lithium triangle.
Non-metallic mining · EnergyDesign of selective collectors and less toxic leaching agents by simulating the mineral–reagent interface at atomic scale. Higher recovery with lower environmental impact.
Large-scale miningMassive search for low-cost OER/HER catalysts for electrolysis. Iridium costs more than gold; finding its replacement defines the economics of green hydrogen in Magallanes and Antofagasta.
Energy · Green H2AI-designed porous materials to desalinate seawater and treat mining effluents with lower energy use. Simulation of ion transport through candidate membranes.
Water resourcesDiscovery of next-generation dielectric coolants — the same case NVIDIA showcased at SC25. If you operate or sell AI infrastructure, this one hits home.
Datacenter · AI infrastructureEmollients, UV filters, formulation stability and ingredient compatibility — one of ALCHEMI's official use cases, with shorter sales cycles than heavy industry.
Retail · Consumer goodsmolecules mapped in half a day — with ALCHEMI, under an hour
«It could drastically improve our molecular property mapping.»
— Qichao Hu, CEO
immersion-fluid candidates evaluated in weeks (plus 100M for OER)
«We hadn't considered searches at the 10–100 million scale; ALCHEMI made extensive sampling surprisingly easy.»
— Takeshi Ibuka, GM AI Innovation
faster conformer search over a space of ~10¹⁰⁰ molecules
«We can completely change the scale and speed of discovery.»
— Julie Brown, EVP & CTO
If they operate at the 10–100 million candidate scale, your first 5,000-candidate campaign runs in an afternoon.
A live, reproducible demo running 100% on-premise on the DGX Spark. No cloud, and not a single molecular structure leaves the room.
We generate or load thousands of candidate structures (electrolytes, catalysts, sorbents) as GPU-resident atomic graphs.
Structural relaxation and molecular dynamics of thousands of systems in parallel with MLIPs (MACE / AIMNet2) on the ALCHEMI Toolkit — near-DFT accuracy at a fraction of the cost.
Ranking by energy, stability and target properties, with live interactive molecular visualization. The client sees their candidate shortlist by the end of the session.
# DGX Spark · GB10 · aarch64 · CUDA 13 $ uv pip install nvalchemi-toolkit $ python run_screening.py --candidates 5000 \ --model mace-mp --stage relax+md [alchemist] loading 5,000 structures → GPU [alchemist] batch MD · 300K · 10ps · NVT ████████████████████░░ 91% ETA 00:04:12 ✓ 5,000 systems relaxed in 38 min (estimated DFT equivalent: ~14 months) ★ top-10 candidates → results/shortlist.zarr $ ▊
Niobium oxide Nb₂O₅ is a promising anode for ultra-fast-charging batteries, but picking the right polymorph and anticipating its lithium intercalation voltage is a classic DFT bottleneck: every structure is hours of quantum compute. We ran that screening with ALCHEMI.
Nb₂O₅ has several polymorphs (T, H, B, TT…) with very similar stabilities but different battery behavior. Ranking them and estimating the lithiation voltage with DFT is expensive and does not scale to dozens of candidates.
We pulled the 9 real Nb₂O₅ polymorphs from Materials Project, relaxed them in a batch with MACE-MP-0 via the ALCHEMI NIM (Batched Geometry Relaxation), ranked them by energy, and estimated the Li voltage profile of the T phase with the same potential — including the metallic-Li reference, computed rather than assumed.
MACE-MP-0 is trained on DFT-PBE, which has known errors for transition-metal oxides (underestimated band gaps, oxidation biases). The relative energies and the voltage profile are useful at the screening level — ranking candidates and reading trends — not as absolute chemical accuracy. We report discrepancies, we do not hide them.
Structures: Materials Project · Compute: ALCHEMI BGR NIM · model MACE-MP-0 (mace_mp_0b2-large) · fmax 0.05 eV/Å · cell relaxed
The same pipeline that ranked the Nb₂O₅ polymorphs answers different industrial questions depending on what you ask it: catalysis, hydrogen, effluents, lead replacement, wear and synthesis routes. Every module runs live from the PoC, with parameters the user picks.
PZT (PbZr₀.₅Ti₀.₅O₃) dominates the piezoelectric market and contains lead. We built its KNN alternative (K₀.₅Na₀.₅NbO₃) by substituting half the K sites in KNbO₃, relaxed both alloys against our own elemental references and compared formation energies per atom.
Second case in the module (wear): NbC reaches B₀ = 333.9 GPa against WC's 496.2 GPa — niobium does not match tungsten in stiffness, and we report that too.
Try it in the PoC ↗
Distance to the convex hull of a full chemical system, with our own energies, consistent among themselves.
⚠ Hull at 0 K: it does not predict which phase wins at process temperature.
Adsorption energy of H, O, OH, H₂O, CO and phenol on low-index surfaces, with ΔG_H* for HER candidates.
⚠ Needs frozen atoms: runs on a local MLIP, not on the NIM.
NVT molecular dynamics at 300–1200 K with a per-structure verdict based on diffusion and preserved local structure.
⚠ The NIM only relaxes: dynamics needs a local MLIP.
Replaces an element in a crystal structure and reports the substitution energy and the local distortion the dopant causes.
⚠ Energies at 0 K without entropy or ZPE: it ranks candidates, it is not absolute accuracy.
Bulk modulus from a Birch-Murnaghan equation of state and the elastic tensor from finite deformations.
⚠ Ideal elasticity at 0 K: no defects, no grain boundaries.
Balances a reaction written as text and computes its ΔE from the relaxed energies of every species.
⚠ Thermodynamics at 0 K: no kinetics, no activation barriers.
The positioning is a funnel: ALCHEMI filters thousands of candidates in minutes, DFT and the lab confirm the few that survive.
Alchemists had the athanor — the furnace where transmutation happened. Ours is an NVIDIA DGX Spark: the Grace Blackwell GB10 superchip with unified memory, in a form factor that installs in the client's office, not in a datacenter.
That means full data sovereignty: formulas, structures and R&D results — the most sensitive asset of a chemical or mining company — never leave the perimeter. And when screening demands more scale, the same pipeline moves unchanged to an H200 cluster or DGX Cloud.
RTX PRO workstation → DGX Spark → H200 cluster / DGX Cloud: same code, different scale.
A 90-minute discovery session with our team: we define your chemical space, run a real screening on the DGX Spark, and you leave with the shortlist.
FREE FOR QUALIFIED COMPANIES · LIMITED SLOTS PER QUARTER