Session T16: Machine Learning and Data in Polymer Physics I

Machine-Learning-Guided Discovery of 19F MRI Agents Enabled by Automated Copolymer Synthesis

Modern polymer science suffers from the curse of multidimensionality. The large chemical space imposed by including combinations of monomers into a statistical copolymer overwhelms polymer synthesis and characterization technology and limits the ability to systematically study structure–property relationships. To tackle this challenge in the context of ¹⁹F magnetic resonance imaging (MRI) agents, we pursued a computer-guided materials discovery approach that combines synergistic innovations in automated flow synthesis and machine learning (ML) method development. A software-controlled, continuous polymer synthesis platform was developed to enable iterative experimental–computational cycles that resulted in the synthesis of 397 unique copolymer compositions within a six-variable compositional space. The nonintuitive design criteria identified by ML, which were accomplished by exploring <0.9% of the overall compositional space, lead to the identification of >10 copolymer compositions that outperformed state-of-the-art materials.   

An interpretable model for polydiketoenamine recyclability

Plastics are versatile and durable yet made for eventual disposal, creating an inevitable environmental issue with plastic waste. Polydiketoenamines (PDKs) offer a solution as novel plastics that are infinitely chemically recyclable through acid-catalyzed hydrolysis. However, experimentally designing new PDK chemistries that target specific properties yet are still chemically recyclable requires significant time and resources. Here, we identify design rules for recyclable PDK chemistries using an interpretable model. We constructed a dataset of molecular features and hydrolysis kinetics by performing high-throughput hybrid-DFT calculations, taking into account the effect of molecular conformation on reaction energetics. We employed a random forest model and symbolic regression to yield an interpretable model based on features from the molecule geometry and electronic structure. This model was validated with 5-fold cross validation, and a subset of data was validated with experimental findings. In collaboration with an experimental team, we validated our model by testing recycling rate of a PDK monomer based on the discovered design rules.    

Measuring mechanical properties at high-throughput using centrifugation

In the era of Big Data and Machine Learning, it is crucial to acquire and catalog enormous datasets of materials properties. Advances in combinatorial chemistry and synthetic biology have drastically increased the synthetic capability for large libraries of materials. However, high-throughput screening remains a challenge particularly for mechanical properties, as these assays often require custom or expensive equipment or do not reach sufficient throughput. In this work, we develop a high-throughput colorimetric screening method for adhesion using a common laboratory centrifuge, multiwell plates, and microparticles. The technique uses centrifugation to apply a homogeneous mechanical detachment force across individual formulations in a multiwell plate. We demonstrate our method using a model pressure sensitive adhesive (PSA) polymer, and we develop a novel high-throughput sample deposition method to prepare films with uniform thickness in each well. After establishing quantitative agreement with a more standard one-at-a-time test method, we demonstrate the throughput of our method, which is limited only by the number of wells in the plates. Further studies explore this centrifugation concept for more general mechanical properties beyond adhesion.   

Benchmarking Machine Learning Models for Polymer Informatics: An Example of Glass Transition Temperature

Various machine learning (ML) models are demonstrated to perform well for polymer's glass transition temperature (Tg) prediction. Nevertheless, they are trained on different datasets, using different structure representations, and based on different feature engineering methods. To provide a fair comparison of different ML techniques and examine the key factors that affect the model performance, we carry out a systematic benchmark study by compiling 79 different ML models and training them on a large and diverse dataset. The three major components in setting up an ML model are structure representations, feature representations, and ML algorithms. In terms of polymer structure representation, we consider the polymer monomer, repeat unit, and oligomer with longer chain structure. Based on that feature representation is calculated, including Morgan fingerprinting with or without substructure frequency, RDKit descriptors, molecular embedding, molecular graph, etc. Afterward, the obtained feature input is trained using different ML algorithms, such as deep neural networks, convolutional neural networks, random forest, support vector machine, LASSO regression, and Gaussian process regression. We evaluate the performance of these ML models using a holdout test set and an extra unlabeled dataset from high-throughput molecular dynamics simulation. The ML model's generalization ability on an unlabeled dataset is especially focused, and the model's sensitivity to topology and the molecular weight of polymers is also taken into consideration. This benchmark study provides not only a guideline for the Tg prediction task, but also a useful reference for other polymer informatics tasks.   

Data-efficient machine learning mimicking human intelligence in fundamental materials science

Machine learning (ML) can be used to solve previously unsolvable materials science problems like highly nonlinear structure-property relationship and high dimensional multifactorial designs. However, requirement for data quality and quantity are not easily satisfied for general ML algorithms. As a result, usually significant experimental resources are needed in generating required data sets. On the other hand, it is widely accepted that human intelligence has a significant advantage in terms of data-efficiency. Experienced materials scientists can usually provide directional insights based on only a few data points. Thus combining the power of machine intelligence (i.e. handling highly nonlinear systems and high dimensional information) and human intelligence (i.e. creating context and making comparison/analogy in materials science language) can enable high data efficiency and reduce development cost for new materials. In this talk, I will give an introduction on data-efficient machine learning that mimics human intelligence in terms of creating digital counterpart of fundamental materials science concepts. These concepts include product by process, materials by construction, graphical representation based pattern recognition, reverse engineering through latent space learning, design by human sensory, fuzzy logic in computational trial and error.   

Withdrawn

Polymer-surface interactions play a significant role in many biological processes and industrial applications. In prior work, machine learning (ML) models have been applied to predict the adhesive free energy of polymer-surface interactions and aid the inverse design. However, in extending these models, one faces the problem that substantially large datasets are not readily available and ML models trained on small datasets have low accuracies.  In this work, we demonstrate a transfer learning (TL) technique with a deep neural network (DNN) to improve the accuracies of ML models trained on small datasets when a larger database from a related system is available. When compared to direct learning (DL), the shared knowledge between the transfer and source tasks improves the performances significantly on small datasets. We explore the limits of database size on accuracy and the optimal tuning of network architecture and parameters for our learning tasks.   

Briding the Scale-Gap: Transfer Learning for Fudamental Polymer Properties using Molecular-Dynamics Simulation Data

Enrichment of databases covering a wide range of molecular structures, materials properties, measurement details, etc. remains as a major challenge in polymer informatics. While molecular dynamics (MD) and other simulations provide a general way to produce arbitrary data with full details of its origin, they often require unrealistically long computational times when macroscopic/product-scale performances must be considered.

In this talk, we will discuss our recent effort to tackle the above challenge by combining MD simulation and transfer learning (TL). As an initial study, MD simulation data is semi-automatically generated for 100+ linear homopolymer melts, which is then utilized as a training set for the “source” task to build a feature encoder for the "target" task. With only 10+ experimental data, TL is able to predict experimental tensile modulus with > 3x higher accuracy compared to the traditional approach that only learns experimental data, demonstrating the usefulness of the proposed approach. This result not only helps apply polymer informatics to real-world problems, but also proposes a way to expand the applicability of theoretical/computational approaches by connecting the scale-gap between the product-scale properties and theoretically accessible quantities.   

Active Learning of Many-Body Transferable Coarse Grained Interactions in Polymers

Despite its widespread use in atomistic modeling, classical molecular dynamics becomes computationally intractable at length and time scales that are of great interest. Coarse graining methods allow fast degrees of freedom to be integrated out of the all-atom system, alleviating both the need to use a small time step and the cost of tracking all degrees of freedom. Bottom-up coarse graining techniques, wherein the thermodynamic properties of the underlying all-atom system are preserved, have seen increased attention with the recent progress in machine learned force fields for ab initio applications. In this work, we extend the idea of Gaussian process based on-the-fly active learning schemes applied to all-atom systems to coarse-grained applications of hydrocarbon liquids. We explore how the inherent interpretability of Gaussian process parameters give novel insight into the learning of coarse grained models, as well as how the active learning framework introduces the possibility of making coarse-grained models more transferable to chemically similar systems.    

Phase Behavior Predictions of Binary Linear Polymer Solutions using Machine Learning

The miscibility of polymers in solvents is important to many polymer processing applications including synthesis, purification, and self-assembly. Predicting the miscibility regions of linear polymer solutions with existing theories has been challenging however, and no universal model exists that quantitatively captures this behavior. Here, we show that a curated experimental data set consisting of 14 linear polymers and 46 solvents can be used to train machine learning models (random forest, XGBoost, and deep neural networks) to predict the cloud point temperature to within 3 °C, and capture various phase behaviors, ranging from upper and lower critical solubility curves, pressure effects (isopleths), and closed-loop behavior. The feature vector used as input is generalizable and consists of a combination of component and state descriptors. Determination of the relative importance of the various descriptors comprising the feature vector for the neural network model is consistent with prior knowledge of polymer phase behavior, including the critical role of polymer size, pressure, and concentration in determining an accurate cloud point temperature.   

Unsupervised learning of sequence-specific aggregation behavior for model copolymers

Sequence specific random block copolymers in dilute conditions exhibit a surprisingly rich phase behavior, including re-entrant phase behavior and large-scale aggregation. We apply a recently developed unsupervised machine learning scheme for local environments [Reinhart, Comput. Mater. Sci., 2021, 196, 110511] to characterize these large-scale, disordered aggregates, which has been shown to be challenging using short-ranged manually derived order parameters. The machine learning algorithm we develop is able to classify the global aggregate structure directly using descriptions of the local environments. We find that the aggregates had overall lower densities than the conventional liquid phases and complex geometries with large interconnected string-like or membrane-like clusters. The resulting characterization provides a deeper understanding of the range of possible self-assembled structures and their relationships to each other.  We demonstrate that by applying unsupervised machine learning to disordered soft matter systems insights can be gained, especially when suitable order parameters are not known.   

Molecular dynamics simulations combined with Gaussian Process regression to investigate block copolymer orientation in thin films

Block copolymer (BCP) thin films are well known to form nanoscale morphologies that are promising as templates and functional materials. The orientation of the morphology is critical for applications and influenced by a host of parameters, including chain architecture, surface tension of each block, and substrate properties. The resulting morphological phase diagram is thus high dimensional and extremely difficult to explore. We deployed molecular dynamics (MD) simulations combined with autonomous experimentation (AE) methods to efficiently explore the vast and complex parameter space of BCP ordering. Our AE pipeline consists of running a MD simulation of BCP chains for a particular point in parameter space, automatically analyzing the results to extract structural metrics, and using an autonomous decision-making algorithm to suggest the next simulation point (the next sets of parameters). This AE loop iterates to build a robust model of the system. The decision-making algorithm is a Gaussian Process regression, customized with a kernel that captures known symmetries of the parameter space.   

Microstructural descriptors for data-driven prediction of energetics and structures of polymer mesophases

Block copolymers are a class of soft-matter systems of particular interest since their composition, size, and architecture can be carefully controlled during synthesis and they can form a wide variety of ordered and disordered mesophases. A quantitative computational framework, self-consistent field theory (SCFT), is also available to assess phase structure and stability, but it is computationally expensive. In this work, we develop a set of sensitive lower-order microstructural descriptors for species density fields, which are generalizations of those previously designed for characterizing two-phase heterogeneous materials and are physically interpretable. Subsequently, we train a theory-embedded machine learning model that incorporates these microstructural descriptors to accurately predict the energetics and structures of copolymer mesophases, using training data generated by SCFT simulations. Because of the inherent translational and rotational invariance of the new microstructural descriptors, our machine learning model achieves perfect global shift-invariance, i.e., the energetics of polymer density fields should be invariant under shifts and rotations. The machine-learned model will help accelerate the discovery of novel functional polymeric materials, benefitting both forward prediction and inverse design.   

Accelerating Langevin Field-Theoretic Simulation with Semantic Segmentation Model

Langevin field-theoretic simulation (L-FTS) can account for the fluctuation effect in a polymer system which is ignored in the self-consistent mean-field theory. Even though L-FTS is computationally efficient compared to traditional particle-based polymer simulations, it requires large computational demand. In order to accelerate the L-FTS, we introduce deep learning (DL). In L-FTS, the functional integral over the pressure field is evaluated using saddle-point approximation whereas the exchange field fluctuates according to the Langevin equation. Using convolutional neural networks for the semantic segmentation task in the computer vision, we directly generate saddle point pressure field for given exchange field. By combining DL and Anderson mixing method, we successfully reduce the number of iterations for finding saddle points, and achieve speedup of 2~3 compared to the Anderson-mixing-only method without sacrifice of accuracy. Our approach is very versatile and efficient enough to be applied to a variety of systems without prior data collection and pre-trained neural network.