Chen, Ji and Xu propose the DiGCA phase classifier, generating Lifshitz-Petrich phase diagrams about two orders of magnitude faster with over 98% classification accuracy
Synopsis
The authors propose a Derivative-informed Graph Convolutional Autoencoder (DiGCA) phase classifier that feeds both the Lifshitz-Petrich model solutions and their derivatives (the nonlocal term G(φ)) into a graph convolutional autoencoder for dimensionality reduction, then classifies with a fully connected neural network, generating phase diagrams over the parameter domain [−0.01,0.05]×[0,1] with over 98% classification accuracy, roughly two orders of magnitude faster than MCMS-RBM, and remaining stable under up to 10% additive white noise.
Figure 1: Order parameters with the corresponding prominent diffraction patterns in the reciprocal space for QC (a), C6(b), LQ(c), T6(d), and Lam(e) states. The top figure also shows the stable (QC)
· Page 4Interpretation
A two-stage DiGCA phase classifier is proposed: the offline stage trains multi-component multi-state graph convolutional autoencoders (MCMS-DiGCA) on solutions together with their derivatives, and the online stage uses a neural network classifier to map encoded solutions to phase diagrams. Compared with the regular graph convolutional autoencoder (GCA-ROM) trained on solutions alone, this method explicitly adds derivatives (the nonlocal term G(φ)) to the training input and trains separate subnetworks for each stable state. The paper provides offline and online algorithmic procedures (Algorithm 1, Algorithm 2) and network schematics, and compares relative L2 errors against the regular GCA-ROM in numerical experiments.
Derivative-informed training substantially improves gradient (nonlocal term) evaluation accuracy while achieving solution prediction performance comparable to the regular method; pointwise relative errors for both the solution and the nonlocal term remain mostly below 5%. The conventional GCA-ROM deviates from the reference solution in phase boundary detection and gives only rough classification trends, whereas DiGCA reconstructs most phase boundaries including their characteristic curvatures and topological connections. Based on relative L2 error comparisons across components over the parameter domain DΞ (Figure 4) and pointwise error maps for the solution and nonlocal term on the test set (Figure 5).
With the deep neural network phase classifier added, the online stage predicts phase diagrams with accuracy exceeding 98%, and the phase diagram changes little under 1%, 5%, and 10% additive white noise. The classifier uses 7 features {ε, α, EQC, EC6, ELQ, ET6, ELam} and outputs a 6-dimensional phase probability, trained with cross-entropy loss and the Adam optimizer for 3000 epochs. The paper reports classification accuracy exceeding 98% and shows phase classification results at different noise levels (Figure 7).
Compared with MCMS-RBM, DiGCA requires less training effort and achieves approximately 100 times faster marginal computation (about two orders of magnitude) at comparable accuracy. MCMS-RBM had already reduced detailed phase diagram generation time from several months to just minutes; DiGCA further lowers training and online computation cost on top of that. Based on cumulative runtime comparisons between MCMS-RBM and MCMS-DiGCA (Figure 8).
Perspective
The framework targets parametric problems whose solutions exhibit multiple states across the parameter domain; this paper validates it on the two-dimensional quasiperiodic Lifshitz-Petrich model over the parameter domain [−0.01,0.05]×[0,1] with six branches (QC, C6, LQ, T6, Lam, Liq), randomly selecting 200 parameter samples per branch and splitting 75% for training and validation. It suits scenarios requiring rapid phase diagram generation such as high-throughput materials discovery and experimental data analysis; the authors also note the mesh can be viewed as a graph structure, that the method applies to unstructured grids as well, and that experimental data could replace high-fidelity solutions to reduce offline cost.
Some minor discrepancies remain near the origin of the phase diagram where all stable states meet; the over 98% classification accuracy and roughly 100-fold speedup come from this paper's numerical experiments, and their behavior on different parameter domains, dimensions, or physical models still needs further verification; the paper mentions that the network can be trained using experimental data, but does not present corresponding experimental validation results.