SynCoTrain: A Dual Classifier PU-learning Framework for Synthesizability Prediction

Technical Summary: SynCoTrain

Problem Statement
The discovery of novel materials is hindered by the inability to reliably predict synthesizability. While thermodynamic stability metrics (e.g., formation energy, distance from the convex hull) offer partial insights, they fail to account for kinetic factors, technological constraints, and specific synthesis conditions that determine whether a material can actually be realized. A critical bottleneck in applying machine learning to this problem is the scarcity of negative data; failed synthesis attempts are rarely published or included in public databases. Consequently, standard supervised learning approaches struggle due to the lack of explicit "unsynthesizable" examples. Furthermore, reliance on a single model introduces inherent architectural biases that can compromise generalizability, particularly when predicting properties for out-of-distribution data.

Methodology
The authors propose SynCoTrain, a semi-supervised machine learning framework designed to predict the synthesizability of oxide crystals. The framework integrates Positive and Unlabeled (PU) Learning with a co-training strategy to address the absence of negative labels and mitigate model bias.

• Co-Training Framework: SynCoTrain employs two distinct Graph Convolutional Neural Networks (GCNNs) as complementary classifiers:

  • ALIGNN (Atomistic Line Graph Neural Network): Encodes atomic bonds and bond angles, reflecting a "chemist's perspective."
  • SchNet: Utilizes continuous convolution filters to encode atomic structures, reflecting a "physicist's perspective."
  • The framework runs two parallel iteration series. In each series, the classifiers alternate roles: one classifier is trained on the current positive set and a sampled subset of unlabeled data (treated as negative), then predicts labels for the remaining unlabeled data. High-confidence predictions are used to expand the positive class for the next iteration, where the other classifier is trained. This iterative exchange of knowledge allows the models to learn from different data views.

• PU Learning Implementation: Within each iteration, the base learners utilize a bagging approach (60 runs) based on the method by Mordelet and Vert. In each run, a random subset of unlabeled data serves as the negative class. The synthesizability score for each data point is the average probability across these 60 runs.

  • Thresholds: A threshold of 0.75 is used to confidently expand the positive class during co-training iterations. A threshold of 0.5 is applied to the final averaged scores to generate binary labels.

• Final Predictor: Once the co-training process concludes (identified as optimal at iteration 2), the generated labels are used to train a final synthesizability predictor. The authors selected SchNet for this final stage, employing regularization techniques to prevent overfitting to the Materials Project data distribution. These techniques include:

  • Label noise injection (flipping 5% of labels).
  • Data augmentation via Gaussian noise perturbation of atomic positions.
  • Weighted loss functions (0.45:0.55 ratio).
  • Dropout layers and Cosine Annealing with Warm Restarts for learning rate scheduling.

Key Results

• Performance Metrics: The model achieved a recall range of 95–97% on internal and leave-out test sets. At the optimal iteration (iteration 2), the model predicted approximately 21% of the unlabeled theoretical data as synthesizable.
• Stability Correlation: The model successfully captured the relationship between stability and synthesizability without explicit training on stability labels. Unstable crystals (energy > 1 eV above the convex hull) were 2.5 times less likely to be classified as synthesizable. However, the model also correctly identified that stability alone is insufficient, as only ~21% of crystals with energy < 1 eV above the hull were classified as synthesizable.
• Generalization: The final predictor was evaluated on external datasets not used in training:
  • OQMD: Showed a higher proportion of synthesizable crystals compared to the Materials Project test set.
  • WBM (Randomly sampled): Showed ambiguous predictions (scores around 0.5), reflecting the artificial nature of the data.
  • iMatGen (Generative): Showed the lowest synthesizability scores, consistent with the generation of unrealistic structures.
• Ground Truth Validation: Using a stability prediction task where ground truth labels were known, the authors demonstrated that the recall values derived from their dynamic and leave-out test sets closely approximated the Ground Truth Recall, validating the reliability of their evaluation metrics.

Significance and Claims
The paper claims that SynCoTrain offers a scalable solution to the synthesizability prediction challenge by leveraging co-training to reduce model bias and handle the lack of negative data. By focusing on oxide crystals—a well-characterized family with extensive experimental data—the authors demonstrate that co-training can achieve high recall while balancing dataset variability and computational efficiency.

The authors position the tool as a resource for the initial stages of high-throughput materials discovery. Rather than claiming to perfectly predict synthesizability, they argue that the model's primary utility lies in filtering out structures unlikely to result in real materials. By potentially eliminating even half of the unsynthesizable candidates, the framework aims to save significant resources in simulation and experimental synthesis attempts. The work highlights the potential of combining existing strong classifiers through co-training to extract more value from available materials data, rather than developing new architectures from scratch.