Faster Chemistry and Materials Discovery with AI-Powered Simulations Using NVIDIA ALCHEMI

Almost all manufactured products are enabled by chemistry and materials science. However, new discoveries are costly and time-consuming and often hindered by trial-and-error approaches. Computational methods such as molecular dynamics with classical force fields traditionally struggle to predict chemical properties and stability accurately.

NVIDIA ALCHEMI (AI Lab for Chemistry and Materials Innovation) enables Chemistry and Materials Science researchers and developers with NVIDIA NIM microservices and toolkits optimized on NVIDIA accelerated computing platforms. This post introduces two new atomistic simulation NIM microservices, the NVIDIA Batched Conformer Search (BCS) NIM and NVIDIA Batched Molecular Dynamics (BMD) NIM.

Accelerated conformer search and molecular dynamics using NVIDIA ALCHEMI

Molecules have multiple three-dimensional shapes, or conformers, each with distinct properties and stabilities. Identifying low-energy conformers is crucial for accurate chemical and material property prediction.

The ALCHEMI BCS NIM efficiently identifies and ranks low-energy conformers of molecules. Traditional quantum chemistry methods for conformer optimization are computationally expensive. BCS NIM uses AIMNet2 as a machine learning interatomic potential (MLIP) to accelerate energy optimization, reducing the time required to generate low-energy conformers compared to quantum chemistry. This is beneficial for high-throughput conformer search campaigns, enabling researchers to rapidly evaluate thermally accessible molecule conformers at near quantum chemistry accuracy.

The BCS NIM works in three steps:

The RDKit or nvMolKit library is used to generate multiple initial conformers from the SMILES format. The maximum number of conformers can be specified or estimated based on the number of rotatable bonds in a molecule.
The conformers are then optimized using AIMNet2 asynchronously into their local energy minima. The optimizations continue until the maximum force on any atom is below 0.005 eV/Å threshold.
The final step involves curating the conformer ensemble by validating bond connectivity based on interatomic distances, discarding high-energy conformers above a user-specified threshold, and deduplicating conformers based on relative differences in interatomic distances.

A bar chart showing that generating more conformers per molecule increases success rates of finding low-energy conformers. — *Figure 1. Percentage of molecules with conformer energies ≤3 kcal/mol, by number of rotatable bonds and number of conformers generated*

Figure 1 shows that as molecular complexity increases—with more rotatable bonds—conformer generation must also increase to ensure sufficient sampling of low-energy conformers. For example, up to 1,000 conformers may be required for a molecule with 12 rotatable bonds to ensure all conformers are within ≤3 kcal/mol of the lowest-energy structure. This shows the need for efficient conformer search methods. With the BCS NIM, average optimization time is 10-100 ms per conformer on an NVIDIA H100 GPU, for F_max = 0.005 eV/Å.

In many workflows, conformer search is combined with molecular dynamics (MD) simulations. The BMD NIM for high-throughput simulations uses MLIPs such as MACE-MPA-0, TensorNet, and AIMNet2 to address this need. Use cases include high-throughput MD simulations of many atomistic systems and agentic AI workflows that require high-throughput asynchronous simulations.

Video 1. Molecular dynamics simulation using the ALCHEMI BMD NIM

Key features of the BMD NIM include:

Dynamic batching: Optimize GPU utilization by dynamically batching atomic systems, for concurrent processing of multiple simulations to maximize throughput.
GPU-based integrators: Perform simulations at a constant number of atoms, volume, and temperature (NVT), or a constant number of atoms, pressure, and temperature (NPT), using a Langevin thermostat and Monte Carlo barostat for temperature and pressure control.
MLIP support: Support for MACE-MPA-0, TensorNet-MatPES-r2SCAN-v2025.1, TensorNet-MatPES-PBE-v2025.1, AIMNet2, AIMNet2-NSE, and AIMNet2-CPCM.

Benchmark charts for molecular dynamics simulation with MACE-MPA-0 and TensorNet MatPES-PBE-v2025.1 by NVIDIA GPU models. — *Figure 2. The ALCHEMI BMD NIM performance benchmark with MACE-MPA-0 and TensorNet-MatPES-PBE-v2025.1*

Figure 2 shows the performance gain (time per atom per step) as batch size increases, demonstrating the benefits of dynamic batching. Improved performance and larger batch size are achieved with newer generations of NVIDIA GPUs, such as the NVIDIA B200. For example, 1.4 μs per atom per step and more than 350,000 atoms simulated on one NVIDIA HGX B200 GPU with TensorNet. The benchmark is performed using batches of randomly sampled structures from the OMat24 dataset.

Accelerating OLED molecular and immersion cooling fluids discovery

NVIDIA customers are achieving groundbreaking results with ALCHEMI’s new capabilities.

Universal Display Corporation (UDC) is accelerating OLED molecular discovery using ALCHEMI BCS and BMD NIM microservices. They’re predicting thermal processing stability up to 10,000x faster than traditional density functional theory (DFT), while achieving near-comparable accuracy. UDC researchers can explore a much larger solution space, increasing the likelihood of finding best-performing OLED molecules.

A diagram illustrating UDC’s OLED two-step virtual screening workflow of using BCS NIM to generate conformers and BMD NIM to predict properties, 10,000X faster than DFT. — *Figure 3. UDC uses NVIDIA ALCHEMI BCS and BMD NIM for OLED discovery*

Starting with a pool of candidate molecules, the researchers use the BCS NIM to identify and rank low-energy conformers. For each conformer, a simulation cell is constructed, and MD simulations are run to assess thermal processing stability. The ability to evaluate thousands of possibilities with unprecedented speed and accuracy was enabled by batching and MLIP. Instead of simulating each molecule separately, ALCHEMI executes numerous simulations simultaneously with the BCS NIM and BMD NIM, accelerating throughput and innovation.

Another collaborator, ENEOS, discovered over 1,000 promising environmentally friendly and non-PFAS immersion cooling fluids using a multi-stage workflow with the BCS NIM, BMD NIM, and the batched DFT microservice. An immersion cooling fluid needs a low dielectric constant to maintain signal integrity and a high flash point to reduce flammability. Starting with 10M molecules, they calculated dipole moments for 1M molecules using the BCS NIM. They used a knowledge-guided pre-training of graph transformer (KPGT) model to predict dipole moments for the remaining 9M, retaining only those with low dipole moments.

Diagram showing ENEOS’s workflow for discovering immersion cooling fluids - candidate molecules are sequentially filtered for low dipole moment, low polarizability, low dielectric constant, and high flash point. — *Figure 4. ENEOS accelerated discovery of immersion cooling fluids with NVIDIA ALCHEMI NIMs* *and microservices*

The batched DFT microservice computed and screened for molecules with low polarizability. Next, the BMD NIM was used to calculate bulk dielectric constants and identify candidates with low dielectric constants. Finally, flash point predictions were made using a KPGT model. This workflow enabled ENEOS to identify over 1,000 promising immersion cooling fluids with both low dielectric constant and high flash point in just three weeks.

Learn more

If you’re at Supercomputing 25, attend the ENEOS talk at the NVIDIA booth and check out the NVIDIA ALCHEMI demo.

Read NVIDIA Accelerated Computing Enables Scientific Breakthroughs for Materials Discovery and Accelerating the Future of Transportation with SES AI’s NVIDIA-Powered Innovation for Electric Vehicles and watch the ALCHEMI video.

Acknowledgments
Thanks to George Fitzgerald, Sean Ryno, Gbolade Kayode, and Darice Liu from Universal Display Corporation, Hideki Ono, Hiroyuki Tsujimoto, Kentaro Yomogita, Yoichiro Kawami, and Masanao Goto from ENEOS, as well as Saee Paliwal and Srimukh Veccham from NVIDIA, for their contributions.