Network architecture¶
Summary¶
==========================================================================
Layer (type:depth-idx) Param #
===========================================================================
Gradient --
├─Sequential: 1-1 --
│ └─ScaleLength: 2-1 --
│ └─AtomRef: 2-2 --
│ └─DistanceAndAngle: 2-3 --
│ └─AtomFeaturizer: 2-4 --
│ │ └─Linear: 3-1 6,080
│ └─EdgeFeaturizer: 2-5 --
│ └─EdgeAdjustor: 2-6 --
│ │ └─Linear: 3-2 192
│ │ └─SiLU: 3-3 --
│ └─ThreeBodyInteration: 2-7 --
│ │ └─NormalizedSphericalBessel: 3-4 --
│ │ └─Linear: 3-5 585
│ │ └─GatedMLP: 3-6 1,152
│ └─M3GNetConv: 2-8 --
│ │ └─GatedMLP: 3-7 33,024
│ │ └─Linear: 3-8 192
│ │ └─GatedMLP: 3-9 33,024
│ │ └─Linear: 3-10 192
│ └─ThreeBodyInteration: 2-9 --
│ │ └─NormalizedSphericalBessel: 3-11 --
│ │ └─Linear: 3-12 585
│ │ └─GatedMLP: 3-13 1,152
│ └─M3GNetConv: 2-10 --
│ │ └─GatedMLP: 3-14 33,024
│ │ └─Linear: 3-15 192
│ │ └─GatedMLP: 3-16 33,024
│ │ └─Linear: 3-17 192
│ └─ThreeBodyInteration: 2-11 --
│ │ └─NormalizedSphericalBessel: 3-18 --
│ │ └─Linear: 3-19 585
│ │ └─GatedMLP: 3-20 1,152
│ └─M3GNetConv: 2-12 --
│ │ └─GatedMLP: 3-21 33,024
│ │ └─Linear: 3-22 192
│ │ └─GatedMLP: 3-23 33,024
│ │ └─Linear: 3-24 192
│ └─AtomWiseReadout: 2-13 --
│ │ └─GatedMLP: 3-25 16,770
===========================================================================
Total params: 227,549
Trainable params: 227,549
Non-trainable params: 0
===========================================================================
Three-body computation¶
triplet_edge_index
Bond featurizer¶
[KME19]
where \(1 \leq m < n_{\max}\)
Atom featurizer¶
Embedding layer
Spherical Bessel function¶
The spherical Bessel function of the first kind
The derivative of spherical Bessel function of the first kind
The spherical Bessel functions \(\{ j_{l}(z_{ln} \frac{r}{r_{c}}) \}_{n=1,\dots}\) form orthogonal basis on \(r \in [0, r_{c}]\),
Here \(z_{ln}\) is the \(n\)th root of \(j_{l}\). We use uses normalized spherical Bessel functions 1
Spherical harmonics with \(m=0\)¶
This definition adopts Condon-Shortley phase.
Legendre polynomial (DLMF 14.10.3)
Derivative of Legendre polynomial
Three-body to bond¶
\(\mathcal{L}_{\sigma}\) is a one-layer perceptron with activation function \(\sigma\).
Cutoff function
Graph Convolution¶
Readout¶
Elemental reference energy¶
We first fit total energies in a training dataset from atom types \(\{ t_{i} \}\). Then, we normalize the residual on the training dataset as
where \(\tilde{E}_{\mathrm{M3GNet}}(\mathbf{A}, \{ \mathbf{r}_{i} \}_{i=1}^{N}, \{ t_{i} \}_{i=1}^{N}) = O(N)\).
Loss function¶
References¶
- JKK19
Ryosuke Jinnouchi, Ferenc Karsai, and Georg Kresse. On-the-fly machine learning force field generation: application to melting points. Phys. Rev. B, 100:014105, Jul 2019. URL: https://link.aps.org/doi/10.1103/PhysRevB.100.014105, doi:10.1103/PhysRevB.100.014105.
- KME19
Emir Kocer, Jeremy K. Mason, and Hakan Erturk. A novel approach to describe chemical environments in high-dimensional neural network potentials. The Journal of Chemical Physics, 150(15):154102, 2019. URL: https://doi.org/10.1063/1.5086167, doi:10.1063/1.5086167.