1. Molecular Representation as Graphs
Traditional molecular descriptors represent molecules as static 2D fingerprinted arrays. To capture physical spatial geometry and quantum mechanical interactions, molecules must be modeled as molecular graphs $G = (V, E)$, where vertices $V$ represent atomic cores and edges $E$ represent chemical bonds. Graph Neural Networks (GNNs) leverage this graph representation to learn node embeddings that capture local chemical environments.
2. Message Passing Neural Networks (MPNN)
The forward pass of an MPNN consists of a message phase and an update phase. During message passing, information is propagated across atomic nodes according to their spatial coordinates and physical properties:
$$m_v^{(t+1)} = sum_{w in N(v)} M_t(h_v^{(t)}, h_w^{(t)}, e_{vw})$$
$$h_v^{(t+1)} = U_t(h_v^{(t)}, m_v^{(t+1)})$$
where $h_v^{(t)}$ is the feature vector of atom $v$ at step $t$, $e_{vw}$ represents bond attributes, and $N(v)$ is the neighborhood of $v$. Equivariant GNNs (EGNNs) enforce translation and rotation invariance, ensuring that rotating a molecule does not alter its predicted quantum energy states.
3. Impact on Materials Informatics
By training these network architectures on quantum chemistry datasets (such as QM9), GNN models can predict DFT-level molecular properties (homo-lumo gaps, binding energies, dipole moments) in milliseconds. This accelerates solid-state battery discovery and polymer design by a factor of $10^5$, establishing a new paradigm in computer-aided materials discovery.