1. Technical Capabilities
A. Architecture & Deployment
- Containerized Microservices: Delivered as Docker containers, enabling rapid deployment on GPU infrastructure.
- RESTful API: Schema-driven HTTP endpoints for seamless integration with laboratory informatics systems (LIMS, ELN, workflow managers).
- High Throughput & Batching: Dynamic batching for efficient GPU utilization, supporting large-scale, concurrent workloads.
- Async/Agentic Workflows: Supports high-volume, asynchronous request patterns for automated pipelines.
- Operational Simplicity: Configurable via environment variables; minimal setup required.
B. Chemistry & Materials Simulation Workflows
- Batched Geometry Relaxation (BGR): High-throughput, batched geometry optimization for molecules and materials.
- Batched Molecular Dynamics (BMD): High-throughput, batched MD simulations, supporting NVE, NVT, and NPT ensembles.
- Periodic & Non-Periodic Systems: Full support for periodic boundary conditions (PBC) and isolated systems.
- Cell Optimization: Variable-cell optimization for periodic systems, with per-request overrides.
- Partial Optimization: Per-atom active masks for selective optimization.
C. Machine Learning Interatomic Potentials (MLIPs)
- Supported Models: MACE, AIMNet2 (multiple variants), TensorNet, and user-supplied custom models.
- DFT-D3(BJ) Dispersion Correction: Optional, for improved van der Waals interactions.
- Implicit Solvation: AIMNet2-CPCM variant for implicit solvent effects.
- External Electric Fields: Can apply during MD simulations.
- Custom DFT-D3 Parameters: User-supplied damping parameters.
D. Data Handling & Integration
- Input Formats: Standard chemistry/materials formats via ASE (CIF, XYZ, extXYZ, etc.).
- Batch Processing: Multiple structures per request (BGR); single system per request (BMD).
- Metadata Support: User-provided metadata fields in requests and replies.
- Rich Output: Optimized structures, MD trajectories, energies, forces, stress, charges, and metadata.
E. Configuration & Customization
- Server Configuration: Model selection, batch size, DFT-D3 enable/params, cell optimization, tolerances, and more via environment variables.
- Per-Request Overrides: Force/pressure tolerances, cell optimization, active masks, etc.
- Automatic Batch Size Estimation: Benchmarks optimal batch size based on GPU and memory.
F. Simulation Control
- Ensembles: NVE, NVT, NPT with Langevin thermostat and Monte Carlo barostat (including anisotropic).
- Restartable Simulations: Long MD runs can be chunked and resumed.
- Advanced Controls: External fields, DFT-D3, implicit solvation, custom models.
2. User Capabilities
A. Workflow Integration
- Python Client Scripts: Provided for both geometry optimization and MD, with async and restart support.
- Command-Line Utilities: For server readiness checks, batch submission, and workflow control.
- Example Datasets: For molecules and materials, periodic and non-periodic, to facilitate onboarding and benchmarking.
B. Informatics Integration
- Schema-Driven Requests/Replies: Detailed control over all simulation parameters and metadata.
- OpenAPI Schema: For programmatic integration and validation.
- Health & Status Endpoints: For monitoring and orchestration in automated pipelines.
C. Usability & Extensibility
- User Metadata: Enables traceability and data provenance.
- Custom Models & Parameters: Extensible to user-supplied MLIPs and simulation settings.
- Batch and Async Processing: Scales to high-throughput screening and automated discovery workflows.
3. Summary Table
| Area | Capabilities |
|---|
| Deployment | Docker containers, GPU-accelerated, ENV VAR config |
| API | RESTful, JSON schema, OpenAPI, health/status/config endpoints |
| Models | MACE, AIMNet2 (variants), TensorNet, custom models |
| Workflows | Batched geometry optimization, batched molecular dynamics |
| Chemistry Support | Periodic/non-periodic, cell optimization, DFT-D3, implicit solvation, external fields |
| Batching | Dynamic, auto-tuned batch size, high throughput |
| User Control | Per-request overrides, active atom masks, metadata, restartable MD |
| Integration | Python scripts, standard file formats, schema-driven, suitable for informatics pipelines |
| Output | Optimized structures, MD trajectories, energies, forces, stress, charges, metadata |
4. Chemistry-Informatics Perspective
- Designed for Integration: RESTful, schema-driven API and standard file formats enable plug-and-play with existing informatics infrastructure.
- High-Throughput Automation: Batch and async processing support large-scale virtual screening and automated discovery.
- Traceability & Data Management: Rich metadata and provenance features align with best practices in scientific data management.
- Extensible & Customizable: Open to new models, simulation settings, and advanced workflows.
5. Benchmarking
tbd
6. References
NVIDIA ALCHEMI BGR and BMD NIM Release 1.4 
Github repository https://github.com/NVIDIA/nvalchemi-toolkit-ops/tree/main