What is MAPLE?
MAPLE (Machine-learning Potential for Landscape Exploration) is a computational chemistry software package designed for molecular modeling and quantum chemical calculations. It provides a unified computational environment suitable for research at different theoretical levels, from small molecules to complex reaction networks.
MAPLE leverages modern machine learning potentials to achieve near-DFT accuracy at a fraction of the computational cost of traditional quantum mechanical methods. Its modular design allows users to perform a wide variety of calculations within a consistent framework, with flexible control over computational parameters and accuracy requirements.
All energy values in MAPLE are expressed in Hartree, and all distances are expressed in Angstroms.
Supported ML Potentials
MAPLE supports a wide range of machine learning interatomic potentials. Each model offers different trade-offs between speed, accuracy, and element coverage.
ANI-2x
General-purpose neural network potential supporting H, C, N, O, S, F, Cl. Trained on DFT data with active learning.
ANI-1x
Neural network potential for organic molecules containing H, C, N, and O. Trained on the ANI-1x data set.
ANI-1ccx
Coupled-cluster accuracy neural network potential for H, C, N, O molecules. Transfer-learned from ANI-1x to CCSD(T)/CBS data.
ANI-1xnr
Non-reactive variant of ANI-1x optimized for stable molecular geometries and vibrational properties.
AIMNet2
Atoms-in-Molecules network potential with broad element coverage and high accuracy for organic and drug-like molecules.
AIMNet2-NSE
AIMNet2 variant with non-stoichiometric energy corrections for improved thermochemical accuracy.
MACE-OFF23 (s/m/l)
Equivariant message-passing potentials for organic molecules. Available in small, medium, and large variants balancing speed and accuracy.
MACE-OMol
MACE potential trained on the OMol25 data set for broad organic molecular coverage with high accuracy.
EGRET
Equivariant graph neural network potential designed for reactive and non-reactive molecular simulations.
UMA
Universal Machine-learning Architecture from Meta. A foundation model covering broad chemical space with a single unified potential.
Computational Capabilities
MAPLE provides a comprehensive suite of computational chemistry methods, all driven by machine learning potentials for fast evaluation on GPU hardware.
Supported Task Types
- Single Point (sp) — Compute molecular energies and properties at fixed geometries.
- Frequency (freq) — Normal mode analysis, RRHO thermochemistry, and IR spectra via mass-weighted Hessian diagonalization.
- Optimization (opt) — Locate equilibrium geometries using L-BFGS, RFO, steepest descent, conjugate gradient, DIIS, or bracketing L-BFGS.
- Transition State (ts) — Saddle point search via NEB, CI-NEB, P-RFO, Dimer, Growing String, or AutoNEB methods.
- PES Scan (scan) — 1D to N-dimensional relaxed or rigid scans along chosen internal coordinates.
- IRC (irc) — Intrinsic reaction coordinate with GS, LQA, HPC, and EulerPC integrators.
Additional Features
- Solvation — GBSA implicit solvation and explicit solvent models with DFT-D4 dispersion corrections.
- Constraints — Fix atoms, bonds, angles, and dihedral angles during optimization and scanning.
- GPU Acceleration — Full CUDA support for ML potential evaluation, providing orders of magnitude speedup over traditional QM methods.
- D4 Dispersion — DFT-D4 dispersion correction available as a global option for any calculation.
Convergence Criteria
MAPLE uses four convergence criteria for geometry optimization and transition state searches. Each criterion can be controlled independently or set via predefined convergence levels.
| Level | f_max (Ha/Bohr) | f_rms (Ha/Bohr) | dp_max (Bohr) | dp_rms (Bohr) |
|---|---|---|---|---|
superloose |
2.5e-01 | 1.7e-01 | 1.0e+00 | 6.7e-01 |
extraloose |
2.5e-02 | 1.7e-02 | 1.0e-01 | 6.7e-02 |
loose |
2.5e-03 | 1.7e-03 | 1.0e-02 | 6.7e-03 |
medium (default) |
4.5e-04 | 3.0e-04 | 1.8e-03 | 1.2e-03 |
tight |
1.5e-05 | 1.0e-05 | 6.0e-05 | 4.0e-05 |
extratight |
1.5e-06 | 1.0e-06 | 6.0e-06 | 4.0e-06 |
Use Cases
MAPLE is well-suited for a variety of computational chemistry workflows:
- Conformational analysis — Rapid geometry optimization and PES scanning of drug-like molecules using ML potentials.
- Reaction mechanism exploration — Transition state search and IRC calculations to map complete reaction pathways.
- Thermochemistry — Frequency calculations with RRHO corrections for free energy and enthalpy estimates.
- High-throughput screening — GPU-accelerated single point calculations across large molecular libraries.
- Benchmarking ML potentials — Compare results from different ML models (ANI, MACE, AIMNet2, UMA) on the same molecular system using a unified interface.