Source-linked AI summary
Polymer Informatics with Multi-Task Learning
Christopher Künneth, Arunkumar Chitteth Rajan, Huan Tran, Lihua Chen, Chiho Kim, Rampi Ramprasad
TL;DR
The paper addresses how correlations among property datasets can be exploited when single-task models provide limited insight into feature–property relationships. It evaluates multi-task learning and finds that NN-MT2 outperforms the other models overall, with multi-task gains especially evident for highly correlated properties and small subclasses.
Problem
Property correlations are potentially exploitable in multi-task learning, while single-task models do not reveal how fingerprint components relate to properties.
Method
The study applies multi-task learning architectures to property prediction and compares them with single-task models based on Gaussian processes and neural networks.
Results
NN-MT2 outperforms the other three models overall, while multi-task models outperform single-task models especially when within-subclass property correlations are high.
Takeaways & Limitations
Multi-task learning successfully improves property prediction by utilizing inherent correlations, particularly when dataset sizes within subclasses are small.
Takeaways & Limitations
The models would benefit from further extension to include uncertainty quantifications for end-users.
Abstract
from arXiv · showhide
Modern data-driven tools are transforming application-specific polymer development cycles. Surrogate models that can be trained to predict the properties of new polymers are becoming commonplace. Nevertheless, these models do not utilize the full breadth of the knowledge available in datasets, which are oftentimes sparse; inherent correlations between different property datasets are disregarded. Here, we demonstrate the potency of multi-task learning approaches that exploit such inherent correlations effectively, particularly when some property dataset sizes are small. Data pertaining to 36 different properties of over $13, 000$ polymers (corresponding to over $23,000$ data points) are coalesced and supplied to deep-learning multi-task architectures. Compared to conventional single-task learning models (that are trained on individual property datasets independently), the multi-task approach is accurate, efficient, scalable, and amenable to transfer learning as more data on the same or different properties become available. Moreover, these models are interpretable. Chemical rules, that explain how certain features control trends in specific property values, emerge from the present work, paving the way for the rational design of application specific polymers meeting desired property or performance objectives.
Results
The dataset contains correlations across 36 polymer properties, and multi-task models generally outperform single-task models, especially for correlated or sparse property groups. NN-MT2 achieves the best overall performance, while fingerprint impacts yield chemical guidelines linking structural features to property trends.
- Correlations in Data: The 36-property dataset contains exploitable positive and negative correlations, including strong relationships among dielectric constants and between thermal, electronic, and transport properties.Examples include positive correlations among frequency-dependent dielectric constants, positive Tg–Tm correlation, and negative band-gap–Tg correlation.
- Single and Multi-Task Models: NN-based single-task models outperform GP-based single-task models, while neural networks also scale more efficiently with dataset, training, and prediction size.The comparison identifies NN models as preferable overall to GP models for this dataset.
- Single and Multi-Task Models: Multi-task models generally outperform single-task models, particularly when property correlations are strong or datasets are sparse.MT models sacrifice some performance on data-rich property groups but improve prediction for correlated and small datasets.
- Single and Multi-Task Models: NN-MT2 delivers the best overall performance across the 36 polymer properties, outperforming NN-MT1, NN-ST, and GP-ST.The concatenation-based NN-MT2 model has the lowest overall error among the four evaluated models.
- Deriving Chemical Guidelines: Fingerprint impacts provide chemical guidelines: atomic rings increase several mechanical, thermal, dielectric, and conductivity-related properties while decreasing several band-gap and stability-related properties.The reported relationships connect ring-induced stiffness, restricted chain motion, and π-electron effects with observed property trends.
Discussion
The study evaluates multi-task learning for polymer property prediction using a coalesced dataset and compares it with single-task approaches. Multi-task models improve performance, efficiency, scalability, and interpretability, especially when correlated property subclasses or small datasets are involved.
- Interpretability: Multi-task learning improves property prediction while using fingerprint impact values to derive chemical guidelines for polymer design.The impact analysis combines SHAP and PCC values and provides an additional knowledge-extraction and validation step.
- Model comparison: NN-MT2 outperforms the other three models overall across the polymer-property benchmark.The benchmark combines 36 properties from over 23,000 data points covering more than 13,000 polymers.
- Correlation-dependent performance: Multi-task models outperform single-task models especially when correlations between properties within a subclass are high.The approach exploits correlations in the coalesced dataset rather than training each property independently.
- Efficiency and scalability: The multi-task approach predicts all properties quickly, avoids training many separate models, and supports scalability and fast retraining as new data or properties become available.These advantages are presented alongside the performance improvements over single-task learning.
- Future work: Uncertainty quantification is identified as a useful future extension for end-users.The discussion notes that the models could be extended further in this direction.
- Scope: The approach is presented as transferable beyond polymers because its learning and fingerprint-impact analyses can be modified for other materials.The authors state that the method is not limited to polymeric materials.
Methods
The study coalesces 36 polymer-property datasets and uses fingerprint-based machine-learning models, including single-task and multi-task architectures, for prediction and interpretation.
- Dataset construction: 36 single-property datasets are fused into a coalesced dataset for holistic multi-task model training.The database contains 36 polymer properties collected from in-house, DFT, experimental, handbook, and online sources.
- Model architectures: Multi-task neural networks use one fingerprint input to predict all properties simultaneously or a selected property iteratively.NN-MT1 uses a multi-head architecture, whereas NN-MT2 concatenates the fingerprint with a 36-component selector vector.
- Polymer representation: Polymer properties are represented with fingerprints spanning atomic, QSPR, and morphological length scales.The fingerprint vector contains 953 components: 371 atomic-scale, 522 QSPR, and 60 morphological descriptors.
- Evaluation: The models are compared using mean squared error across five-fold cross-validation, with means and 68% confidence intervals reported.The same loss function is used for the four machine-learning models and five training datasets.
- Baseline models: Gaussian-process regression provides uncertainty estimates but becomes unsuitable for big-data learning because kernel-matrix inversion scales quadratically with feature count.Prediction means are derived from probability distributions, while their confidence intervals define uncertainty.
- Interpretability: SHAP values are combined with their correlation to fingerprint components to estimate signed fingerprint impact on polymer-property predictions.Mean absolute SHAP values measure total impact, while the correlation sign captures positive or negative proportionality.