Atomistic Line Graph Neural Network (ALIGNN)

• Atomistic Line Graph Neural Network (ALIGNN) is a GNN that explicitly integrates bond lengths and bond angles via atomistic and line graph representations.
• It utilizes an edge-gated graph convolution mechanism to propagate both atomic and angular information, enhancing the capture of many-body geometric effects.
• Benchmark studies show that ALIGNN outperforms traditional models on materials datasets, making it a versatile tool for scalar property prediction and spectral analysis.

to structure article. Because user asked encyclopedia article. Need maybe include formulas for graph update, line graph update, loss MSE for spectra? Could include only if essential. Since target audience technical. But avoid too many. Maybe include core ALIGNN equations from original and one or two application equations. Yet article about ALIGNN, not each application's domain formula. Better include core architecture equations from (Choudhary et al., 2021) and generic update from later. Maybe mention DPA3 scaling law in context section but that's not directly ALIGNN; still relevant as line-graph family. But user asked ALIGNN topic, comprehensive article. Could mention ALIGNN-d and DPA3 as descendants/extensions. Must ensure claim present in data. DPA3 section has many details; can mention DPA3 treats ALIGNN as one of earliest and best-known line-graph GNNs and generalizes to line graph series, with experiments finding K=2 best (Zhang et al., 2 Jun 2025). That's allowed.

Need search tool. Let's do arXiv search for ALIGNN maybe original and recent. Atomistic Line Graph Neural Network (ALIGNN) is a graph neural network architecture for atomistic systems that explicitly incorporates both interatomic distances and bond angles by performing message passing on an atomistic bond graph and on its line graph, whose nodes correspond to bonds and whose edges correspond to bond-angle triplets. Introduced for crystals and molecules, ALIGNN was evaluated on 52 solid-state and molecular properties from the JARVIS-DFT, Materials Project, and QM9 databases, and was reported to outperform some previously reported GNN models on atomistic prediction tasks by up to 85% in accuracy with better or comparable model training speed (Choudhary et al., 2021).

1. Origin and design objective

ALIGNN was proposed in response to a specific limitation of distance-centered atomistic GNNs. Conventional models such as CGCNN, SchNet, and MEGNet represent a structure as atoms connected by bonds or neighbor relations, with edge attributes dominated by interatomic distances. In that setting, many-body geometric effects are only implicit. The original ALIGNN paper emphasized that many solid-state and molecular properties are highly sensitive to local geometric details, especially bond angles and distortions of coordination polyhedra, and therefore introduced an explicit angular pathway through a line graph (Choudhary et al., 2021).

The original benchmark program was deliberately broad. ALIGNN was trained on the Materials Project 2018.6.1 dataset with approximately 69,239 materials for formation energy and band gap, on JARVIS-DFT version 2021.8.18 with 55,722 materials and 29 regression targets, and on QM9 with 130,829 molecules and 11 molecular properties (Choudhary et al., 2021). This breadth established ALIGNN not as a domain-specific architecture, but as a general-purpose graph formalism for atomistic learning.

Subsequent work treated ALIGNN as a canonical representative of the “graph + line graph” paradigm. In the DeepCrysTet comparison study, ALIGNN was described as “currently among the state-of-the-art models with the best performance on the Materials Project dataset” because it explicitly encodes bond-angle information (Tsuruta et al., 2023). Later line-graph architectures, including ALIGNN-d and DPA3, were framed as extensions of this same idea rather than departures from it (Hsu et al., 2021, Zhang et al., 2 Jun 2025).

2. Atomistic graph, line graph, and input descriptors

For crystals, ALIGNN constructs a periodic 12-nearest-neighbor graph and then expands the neighbor list to include all atoms within the neighbor shell of the 12th-nearest neighbor. The resulting graph $G=(v,\mathcal{E})$ has atoms as nodes and neighbor pairs as edges. For molecules, the same conceptual graph is used with a shorter radial cutoff, up to $5$ Å instead of $8$ Å (Choudhary et al., 2021).

The atomistic line graph $L(g)$ is built from that bond graph. Each node in the line graph corresponds to an edge $(i,j)$ in the atomistic graph, and each edge in the line graph connects two bonds that share a common atom, such as $(i,j)$ and $(j,k)$. That line-graph edge represents the bond angle $\theta_{ijk}$ at the central atom $j$. The construction converts three-body geometry into an ordinary graph-neighborhood relation, allowing angle-sensitive message passing without leaving the graph-convolutional framework (Choudhary et al., 2021).

The original ALIGNN implementation uses atomic features based on elemental properties and geometric features based on radial basis expansions. Each node has 9 input features based on atomic species, including electronegativity, group number, covalent radius, valence electrons, first ionization energy, electron affinity, block, and atomic volume. Bond features are radial-basis expansions of interatomic distance, with 80 distance RBFs by default. Triplet features are radial-basis expansions of bond angle, with 40 angle RBFs by default. The bond angle is defined by

$$\theta_{ijk}=\arccos\frac{\mathbf{r}_{ij}\cdot\mathbf{r}_{jk}}{\|\mathbf{r}_{ij}\|\,\|\mathbf{r}_{jk}\|}.$$

These atom, bond, and angle features are embedded before message passing (Choudhary et al., 2021).

Later applications largely retained this representation. RamanGPT, for example, used the standard ALIGNN architecture with 92 atomic input features, a cutoff radius of 8.0 Å, and a maximum of 12 neighbors per atom; line-graph edges encoded bond-angle triplets for Raman-spectrum prediction (Abel et al., 2 Jun 2026). The persistence of this featurization across electronic, vibrational, optical, adsorption, and force-field tasks indicates that ALIGNN’s representation is unusually reusable across property classes.

3. Edge-gated graph–line-graph message passing

The core computational unit in ALIGNN is the edge-gated graph convolution, applied both to the atomistic graph and to the line graph. On a generic graph with node features $5$0 and edge features $5$1, the original paper writes the node and edge updates as

$5$2

$5$3

with SiLU activations and normalization layers (Choudhary et al., 2021).

An ALIGNN layer couples two such convolutions: $5$4

$5$5

The first operation updates bond and triplet features on the line graph; the second uses the updated bond messages $5$6 as edge inputs on the atomistic graph to update atom and bond embeddings. Angular information therefore reaches atom representations at every ALIGNN layer rather than only through deep implicit composition (Choudhary et al., 2021).

The default configuration in the original work used $5$7 ALIGNN layers followed by $5$8 additional edge-gated graph-convolution layers on the bond graph, with embedding size 64, hidden feature size 256, batch normalization, SiLU activation, and global average pooling over atoms before the final readout (Choudhary et al., 2021). This architecture remained close to the official implementation reused in later studies, including DeepCrysTet baselines, RamanGPT, and defect modeling (Tsuruta et al., 2023, Abel et al., 2 Jun 2026, Rahman et al., 2023).

Architecturally, ALIGNN occupies a middle position between pairwise crystal GNNs and fully directional or equivariant models. It remains a scalar graph network, but it introduces explicit three-body structure through a separate graph object. This design choice proved sufficiently expressive for scalar properties, spectra, and potential-energy learning across a wide range of chemistries.

4. Benchmark performance and empirical behavior

The original ALIGNN benchmarks showed systematic gains over descriptor-based and graph-based baselines. On Materials Project, ALIGNN reached a test MAE of 0.022 eV/atom for formation energy and 0.218 eV for band gap. On JARVIS-DFT, selected test MAEs included 0.033 eV/atom for formation energy, 0.076 eV for energy above hull, 10.4 GPa for bulk modulus, 9.48 GPa for shear modulus, and 0.26 $5$9 for magnetic moment. On QM9, ALIGNN achieved 0.0214 eV for HOMO, 0.0195 eV for LUMO, and 0.0146 D for dipole moment (Choudhary et al., 2021).

Benchmark	ALIGNN result	Comparison in the source
MP formation energy	0.022 eV/atom	Better than CGCNN 0.039, MEGNet 0.028, SchNet 0.035
MP band gap	0.218 eV	Better than CGCNN 0.388 and MEGNet 0.33
JARVIS formation energy	0.033 eV/atom	Better than CFID 0.14 and CGCNN 0.063
JARVIS $8$0	0.076 eV	Better than CFID 0.22 and CGCNN 0.17
QM9 dipole moment	0.0146 D	Better than SchNet 0.033, MEGNet 0.05, DimeNet++ 0.0297

The benchmark program also included classification tasks derived from JARVIS-DFT. Reported ROC AUCs were 0.94 for stable versus unstable, 0.92 for metal versus non-metal, 0.91 for magnetic versus non-magnetic, and 0.92 for high versus low p-type Seebeck coefficient (Choudhary et al., 2021). This established ALIGNN as a usable screening model as well as a regressor.

An ablation study in the original work clarified where the performance gain comes from. GCN-only models saturated at approximately 44 meV/atom on JARVIS-DFT formation energy, whereas ALIGNN-only models saturated at approximately 34 meV/atom, a 29.14% reduction. Mixed ALIGNN+GCN stacks achieved the best trade-offs. The source interpreted this as direct evidence that explicit propagation of bond-angle information improves generalization for properties tied to local bonding geometry (Choudhary et al., 2021).

5. Extensions across spectra, force fields, defects, adsorption, and discovery

ALIGNN was rapidly extended from scalar property prediction to structured outputs and simulation surrogates.

Domain	Use of ALIGNN	Reported result
Electronic DOS	300-bin DOS prediction	D-ALIGNN average MAE 0.009 states/eV/electron
Phonons	Phonon DOS and derived properties	78% of test samples have MAE < 0.086; stability accuracy ~90%
Universal force field	ALIGNN-FF on 89 elements	0.086 eV/atom energy MAE and 0.047 eV/Å force MAE
Defects	Defect formation energy prediction	RMSE around 0.3 eV
2D magnets	Magnetic classification and $8$1	93% accuracy and 0.039 eV/atom MAE
NLO	Kurtz–Perry coefficient prediction	82.5% high-confidence predictions
Raman	200-bin Raman spectra	Mean MAE 0.032; 42.5% with cosine similarity $8$2

For spectral learning, ALIGNN proved effective on both electronic and vibrational densities of states. In electronic DOS prediction, D-ALIGNN directly predicted a 300-dimensional DOS vector over $8$3 eV and achieved an average MAE of 0.009 states/eV/electron, versus a baseline MAE of 0.031 states/eV/electron (Kaundinya et al., 2022). In phonon DOS prediction, ALIGNN was trained on over 14,000 phonon spectra from JARVIS-DFT; 78% of test samples had MAE < 0.086, dynamic-stability classification reached about 90% accuracy, and DOS-mediated predictions of $8$4, $8$5, and $8$6 outperformed both direct scalar models and Debye or Born–von Kármán approximations (Gurunathan et al., 2022).

The same graph formalism was then used for optical and spectroscopic targets. A study of optical absorption spectra for 2,681 oxides, chalcogenides, and related compounds trained ALIGNN to predict $8$7 on a 1501-point grid and then used first-layer ALIGNN embeddings for hierarchical clustering; 75% of materials had MAE < 0.14 in log-absorption space, and the learned feature space grouped materials by both spectral and chemical similarity (Takahashi et al., 20 Oct 2025). A dielectric-function model trained two independent ALIGNN networks on 6,918 TB-mBJ dielectric functions from JARVIS-DFT and, when coupled to JARVIS-tools SLME postprocessing, achieved an SLME MAE of 1.952 percentage points with $8$8 on the test set (Ginter et al., 9 Oct 2025). RamanGPT used a standard ALIGNN forward model trained on 5,099 DFPT Raman spectra; on the 509-material test set it reported mean MAE 0.032, median MAE 0.029, and 42.5% of test spectra with cosine similarity $8$9, a threshold the authors treated as “reasonable” qualitative matching (Abel et al., 2 Jun 2026).

ALIGNN also became a basis for universal and transferable force fields. ALIGNN-FF, trained on 307,113 configurations and 9,593,385 force entries drawn from JARVIS-DFT, used a combined energy–force loss and predicted forces by automatic differentiation of the learned energy. For the chosen force weighting $L(g)$0, the reported test errors were 0.086 eV/atom for energy and 0.047 eV/Å for forces. When used to relax 23,495 JARVIS solids with $L(g)$1 atoms per cell, lattice-constant MAEs were 0.11 Å, 0.11 Å, and 0.13 Å for $L(g)$2, $L(g)$3, and $L(g)$4, with 0.08 eV/atom formation-energy MAE (Choudhary et al., 2022). A later high-entropy-alloy study used the pre-trained ALIGNN-FF as a frozen surrogate, reported approximately $L(g)$5 lower computational cost than DFT for energy evaluation, and achieved 82% accuracy in predicting experimentally reported FCC or BCC structures across 132 single-phase HEAs (Beaver et al., 2024).

Defect energetics provided another early test of ALIGNN beyond pristine crystals. In a large semiconductor-defect dataset covering vacancies, interstitials, antisites, impurities, and complexes, ALIGNN outperformed CGCNN and MEGNET and produced defect-formation-energy RMSEs around 0.3 eV, described as 98% accuracy relative to the dataset range. The same work used ALIGNN energies in a gradient-free structural search and reduced charge-transition-level RMSE from 0.37 eV for unoptimized structures to 0.17 eV after ALIGNN-based optimization (Rahman et al., 2023).

High-throughput discovery studies then adopted ALIGNN as a screening engine. In 2D magnet discovery, two ALIGNN models were trained on 1,190 magnetic monolayers, yielding approximately 93% accuracy for magnetic classification and approximately 0.039 eV/atom MAE for $L(g)$6; the pipeline filtered 11,100 CDVAE-generated candidates and ultimately identified 167 magnetic monolayers with $L(g)$7 eV/atom and total magnetization $L(g)$8 after DFT validation (Elrashidy et al., 2023). In nonlinear optics, an ALIGNN model trained on 2,274 non-centrosymmetric crystals from NOEMD predicted the Kurtz–Perry coefficient $L(g)$9 and achieved 82.5% high-confidence predictions under the criterion $(i,j)$0 pm/V or $(i,j)$1 (Alkabakibi et al., 28 Apr 2025). In metal–organic frameworks, ALIGNN was trained on 137,953 hMOFs and reported a multi-output CO$(i,j)$2 adsorption MAE of 0.19 mol kg$(i,j)$3; on QMOF band gaps it achieved 0.20 eV MAE, improving on the cited CGCNN value of 0.274 eV (Choudhary et al., 2021).

Taken together, these studies show that ALIGNN is not confined to equilibrium scalar regression. It has functioned as a predictor of spectra, a surrogate potential, a defect-energy model, and a screening model for adsorption, magnetism, and nonlinear optics.

6. Limitations, critiques, and line-graph descendants

The clearest critique of ALIGNN has come from work arguing that graph and line-graph representations still discard essential three-dimensional information. DeepCrysTet, which represents crystals by a Delaunay tetrahedral mesh, used ALIGNN as its strongest graph-based baseline on Materials Project version 2018.10.18. In that study, ALIGNN reached 75.6% accuracy for 7-way crystal-system classification and 64.3% for 230-way space-group classification, whereas DeepCrysTet reached 97.5% and 90.3%. At the same time, ALIGNN remained best on formation energy, with MAE 0.024 eV/atom versus 0.062 for DeepCrysTet, and on band gap, with 0.211 eV versus 0.279. The authors concluded that bond-angle-enriched graphs improve local 3D information but still struggle with full cell geometry and global symmetry, especially for triclinic and monoclinic systems (Tsuruta et al., 2023).

Other limitations are architectural and computational. The original ALIGNN paper noted that adding line-graph message passing roughly doubles the cost of a GCN layer, and deeper networks show saturating performance and potential degradation beyond about 4–6 layers (Choudhary et al., 2021). Application papers likewise identify regime-specific failure modes: Raman prediction degrades for materials with numerous sharp and densely packed peaks and for compounds with substantial activity beyond the 1000 cm$(i,j)$4 training window (Abel et al., 2 Jun 2026), while MOF adsorption models degrade for high-void-fraction frameworks and long-range pore descriptors such as LCD and PLD (Choudhary et al., 2021).

Two major descendants address these issues by extending the same representation rather than discarding it. ALIGNN-d augments ALIGNN with dihedral angles. On dynamically disordered Cu(II) aqua complexes, ALIGNN-d used 54, 80, and 111 edges for 4-, 5-, and 6-coordinate cases, compared with 78, 120, and 171 for the fully connected $(i,j)$5 representation, i.e. 31–35% fewer edges, while achieving validation loss essentially identical to $(i,j)$6 (Hsu et al., 2021). DPA3 generalizes the line-graph idea to a line graph series and explicitly treats ALIGNN as one of the earliest and best-known line-graph GNNs. Its experiments found that moving from order $(i,j)$7 to $(i,j)$8 gives a large improvement, but $(i,j)$9 hurts accuracy and training stability, so the best default remains the atom-plus-bonds/angles regime already used by ALIGNN (Zhang et al., 2 Jun 2025).

A plausible implication is that ALIGNN’s most durable contribution is not a single benchmark number but a representational principle: local atomistic learning benefits markedly from elevating bonds and bond angles to first-class graph objects. Later work has extended that principle toward dihedrals, scaling-law studies, and large atomistic models, but the original atomistic graph plus one line graph remains a strong and widely reused operating point.