Session W16: Machine Learning and Data in Polymer Physics II

Soft, biologically inspired materials for neuromorphic memristors and memcapacitors

This talk will describe the development and characterization of neuromorphic elements for brain-inspired computing such as memristors and memcapacitors assembled from soft materials like lipids, polymers, peptides, proteins, and redox-active molecules. These materials do not function like true biological synapses, but they do exhibit highly nonlinear dynamical properties reminiscent of action potential propagation in neural networks in the central nervous system and use very similar sets of dynamical equations. Currently, the fragility and short lifetimes of soft-matter-based membranes such as lipid bilayers limit the assembly of individual biomolecular memristors or memcapacitors into extended biomembrane-based neuromorphic networks capable of brain-like sensing or computation. We use the droplet interface bilayer (DIB) platform to assemble and electrochemically characterize model bilayer membranes consisting of either amphiphilic polymer or lipid molecules that exhibit complex, nonlinear behaviors such as those found in biology, but that are also stable enough to be configurable into massively parallel neuromorphic circuits. These efforts are informed by vibrational sum-frequency generation spectroscopy, small-angle neutron scattering, fluorescence microscopy, and molecular dynamics simulations at both air-liquid and liquid-liquid interfaces. These experiments reveal previously hidden yet critically important structural and dynamical interactions occurring in both the headgroup region and in the hydrophobic tails that impact not only the memristive and memcapacitive behaviors but also membrane stability.   

Machine learning approach to identify critical configurations for strong electronic coupling

In an organic semiconductor material, the rate of interchain charge transfer depends on the electronic coupling parameter between neighboring monomers. The rate of charge transfer depends on not only the highest coupling but also the distribution of it. Here, we develop an automated method to calculate electronic coupling that considers the effect of polarization and hence gives a more accurate description. The automatic process allows us to explore various configurations and generate data to train and develop a model using machine learning approaches. Here we have examined the regioregular Poly(3-hexylthiophene) P3HT (donor) as an example. We explore the real structure of P3HT using molecular dynamics simulation. The neighboring P3HT monomer orientations are fed to the machine learning model to obtain electronic coupling distribution. The electronic coupling for an amorphous P3HT system exhibit a skewed distribution with rare but strong coupling above 100 meV. The probability of finding substantial coupling states increases as we anneal the system. The method has also allowed us to obtain the possible key configurations with high electronic coupling.   

The use of small angle neutron scattering data to support dark field data analysis in far-field interferometry

Small angle neutron scattering (SANS) is a useful measurement for probing bulk structure on the order of 1 nm to 10 μm in soft matter ranging from biological materials and colloids to polymer films and more.  One limitation is that it provides a sample-averaged view of the structure, making the study of heterogeneous materials difficult with SANS alone. However, a new type of grating-based far field interferometer currently under development at the National Institute of Standards and Technology (NIST) will enable the collection of spatially resolved (~100 μm) structural information on the same length scales as SANS. The dark field intensity collected with this technique includes the relevant structural information and is related to SANS through a Hankel transform, and so in this talk, we will discuss how our knowledge of structural systems in the SANS (Fourier) space can be used to better understand how these features present in the dark field (correlation) space. This is not only useful for experimental planning and understanding the use cases and limitations of this technique, but will also aid development of the high-throughput data handling methods (~TB/day) required for this instrument.   

High-throughput microrheology of polymer solutions and gels

Passive probe microrheology has become a popular method for characterizing viscoelasticity on small fluid samples, and holds significant potential for informing rheological design over a wide formulation space with limited material. Realizing this potential will require automated, high throughput data acquisition and analysis. Here, we report a new method for extracting microrheology information using differential dynamic microscopy (DDM). Using Fourier-domain analysis of video images, DDM can extract the mean-squared displacement in systems that would otherwise be difficult to measure using conventional particle tracking. Combining DDM with downsampling by Gaussian process regression, we demonstrate that DDM microrheology can be performed in real time. This rapid acceleration is leveraged to integrate fully automated sample preparation, data acquisition and analysis to demonstrate autonomous, high-throughput microrheology characterization. We illustrate the utility of high-throughput microrheology through two examples – in situ characterization of viscosity during polyelectrolyte coacervation, and kinetic profiling of gelation in protein solutions. The results highlight the considerable promise of automated microrheology to aid in the design of polymeric fluids and soft solids.   

Predicting the Glass Transition of Complex Polymers via Integration of Machine Learning, Theory and Molecular Simulations

Semiconducting conjugated polymers (CPs) are attractive organic electronic materials for a wide range of applications due to their unique properties such as easy processability, tunable electrical performance, and mechanical flexibility. Despite tremendous efforts, design and prediction of  T_g remain notably challenging for CPs due to their complex chain architecture associated with diverse chemical building blocks. In this work, we establish an integrated framework based on machine learning (ML) and molecular simulations to predict T_g for a diverse set of CPs and other polymers with a drastic difference in their chemical structures. Informed from informatics and molecular theory, the developed ML model takes the geometry of diverse chemical building blocks to define simplified structural features to make T_g prediction, which is further validated by experimental measurement. Moreover, the use of molecular simulation and theory in conjunction with ML uncovers the critical roles of key molecular features in influencing the glass transition temperature as well as dynamics heterogeneity associated with glass formation of complex polymers. The established predictive framework and ML model could be ready to use for the design of high-performance CPs and relevant materials via molecular engineering.   

Diverse property-spectrum of flavors of polyolefins: A data analysis study

It is known that synthesis and processing conditions drastically influence the morphology of a polymer resulting in many flavours of the same polymer displaying wide variations of a given property value. One such feature that leads to property variation for the same polymer is the molecular weight. Variations in molecular weight lead to many flavors of the homopolymer polyethylene with glass transition temperatures ranging from ~150-350 K for neat resins. In this regard, the numerical data of the empirically known flavors of polyolefins are examined for their property variations through a comprehensive data analysis. These polymers involve hundreds of flavors of polyethylene and polypropylene resins whose thermal, rheological, thermodynamic and mechanical properties are analyzed for understanding their property variations. In addition, statistical correlations between the properties are quantified by relating them to features such as density, molecular weight, and crystallinity. Such data analysis studies are useful (1) to obtain a comprehensive understanding of the sources of property value variations, and (2) improving existing machine learning models for property prediction by incorporating features in the fingerprint (i.e., representation of polymers) derived from key features known to affect properties.   

Machine Learning Discovery of Multi-Functional Polyimides

Polyimides have been widely used in modern industries but it takes decades of experimental efforts to develop a successful product. Aiming to discover high-performance polyimides, we utilize computational methods of machine learning (ML) techniques and molecular dynamics (MD) simulations. We first build a dataset of more than 8 million hypothetical polyimides based on the polycondensation of existing dianhydride and diamine/diisocyanate molecules. Then we establish multiple ML models for thermal and mechanical properties of polyimides based on their experimentally reported values. The obtained ML models demonstrate excellent predictive performance and identify the key chemical substructures influencing the thermal and mechanical properties of polyimides. Applying the well-trained ML models, we obtain property predictions of the 8 million hypothetical polyimides. In such way, we explore the whole hypothetical dataset and identify 3 novel polyimides that have better combined properties than existing ones through Pareto frontier analysis. Furthermore, we validate the ML predictions through all-atom MD simulations and examine their experimental synthesizability. This study discovers novel polyimides and guides the further experimental synthesis of innovative polyimides.   

HI vs AI: designing solvent-free brush networks with tissue-like mechanical properties

The ability to synthesize elastomeric and tissue-mimetic materials with programmable mechanical properties, enabling robust performance in constantly evolving conditions, is essential for future progress in soft materials design. We developed a design strategy of solvent-free networks made by crosslinking brush-like strands which is based on a theoretical model of the nonlinear network deformation and Machine Learning approach. Implementing a multi-layer feedforward artificial neural network (ANN), we take advantage of model-predicted structure-property cross-correlations between system code parameters A_i(l,v,b,ρ) and B_j(n_sc,n_g,n_x), describing system-specific chemistry and network strand architecture, and equilibrium mechanical properties of networks defined by the structural shear modulus G and firmness β. The ANN was trained by minimizing the mean-square error with Bayesian regularization to avoid overfitting for a data set of 118 experimental stress-deformation curves of networks with brush-like strands of PBA and PDMS having structural modulus G < 90 kPa and 0.01 ≤ β ≤ 0.9. The trained ANN was capable of predicting network mechanical properties based on strand chemistry and architecture with better than 95% confidence.   

Illuminating Stress and Failure in Polyethylene with a Neural Network Potential

Polyethylene is a material with a range of applications which depend less on its intrinsic chemistry and more upon its structure on the nano- to micro-meter scale.  Low-density polyethylene, with its more open, disordered structure and yields softer, more ductile materials.  High-density and ultra-high molecular weight polyethylene (UHMWPE) can be used to make materials strong enough to for body armor and soldier protection.  It is therefore paramount to understand how the mechanical properties of UHMWPE depend on its microstructure and how defects can impair the strength and lead to material failure.  This is a difficult task however, given that we require a potential which can model bond breaking while being efficient and accurate enough to simulate bulk PE over nanosecond timescales.  We have thus trained a neural network potential (NNP) based on the SCAN density functional potential energy surface which yields energies within 1 meV/atom of the ab-initio results yet is several orders of magnitude more efficient.  We have used this potential to study the mechanical properties and failure of single PE strands and a PE crystal both for pristine samples and in the presence of defects, such as a single PE knot.  We find that failure occurs depend on bond length deviations from the instantaneous average, not necessarily on the bond length alone, and that failure in knotted PE occurs at the knot entrance, similar to previous results by Klein and Saitta.  Our results shed light on failure mechanisms in UHMWPE, how they are related to the material microstructure, and how they might be mitigated to yield improved properties.   

Machine Learning-based Study of Mechanical Properties of Dynamically Crosslinked Polymer Networks

In recent years, polymer composites synthesized from dynamically crosslinked networks have demonstrated improved and innovative properties including shape memory, adhesive, self-healing and malleability. Such polymer networks consist of multiple components that can interact with each other in complex ways. The polymer backbone and the crosslinking agents may be designed in a variety of architectures with the potential to deliver a wide range of physical properties. Understanding the effect of polymer network architecture on the resulting properties of the material is an important and challenging task. Computational models and Machine Learning techniques can provide a useful platform to investigate structure-property relations of crosslinked polymer networks. 

In this study, we utilize Molecular Dynamics (MD) simulations to investigate the relationship between polymer network configuration and the resulting mechanical properties of crosslinked polymer composites. MD simulations are employed to generate stress-strain curves for a variety of crosslinker and backbone polymer configurations. The results of the MD simulations are gathered as the reference data set to be utilized within a Machine Learning (ML) framework. We establish “3D images” of the polymer network configurations obtained from MD simulations and build Convolutional Neural Networks (CNNs) in order to investigate the relationship between the architecture of the network and the mechanical behavior of the material. We discuss the efficiency and accuracy of the CNN in evaluating the mechanical properties of each system and study the impact of the configurational details of each network under set initial conditions. The results of this work provide new insight into the complex architecture of crosslinked polymer networks and help identify material structures that can deliver desired mechanical properties.    

Application of machine-learned constitutive relations for well-entangled polymer melt flows

We investigated the effectiveness of a Machine Learning (ML) approach to learn the constitutive relation of well-entangled polymer melts. In particular, we use a Gaussian Process regression on training data (e.g., stress, strain rate, number of entanglements) obtained from microscopic Slip-link simulation under various flow conditions. The learned constitutive relations (given in differential form) were then used within macroscopic Smooth Particle Hydrodynamics simulations of standard flows (i.e., parallel plates and contraction expansion).

Compared to currently used hierarchical simulation methods, which directly couple the microscopic polymer dynamics and the macroscopic flows, our method is ~100 times faster and drastically reduces the statistical fluctuations in the flow predictions, while maintaining a clear connection to the underlying molecular model. In this presentation, we will use benchmark Multi-Scale simulation results to present a detailed error analysis of our ML predictions and their computational efficiency. 

This work was partially supported by Japan Society for the Promotion of Science Kakenhi (19H01862, 20K037865), SPIRITS 2020 of Kyoto University, Ogasawara Foundation.   

Machine Learning Parametrization of a Coarse-grained Epoxy Model at Varying Crosslink Density

A persistent challenge in molecular modeling of thermoset polymers is capturing the effects of chemical composition and degree of crosslinking (DC) on dynamical and mechanical properties with high computational efficiency. We established a coarse-graining (CG) approach combining the energy renormalization method with Gaussian process surrogate models of molecular dynamics simulations. This allows a machine-learning informed functional calibration of DC-dependent CG force field parameters. Taking versatile epoxy resins consisting of Bisphenol A diglycidyl ether combined with curing agent of either 4,4-Diaminodicyclohexylmethane or polyoxypropylene diamines, we demonstrated excellent agreement between all-atom and CG predictions for density, Debye-Waller factor, Young's modulus, and yield stress at any DC. We further introduced a surrogate model-enabled simplification of the functional forms of 14 nonbonded calibration parameters by quantifying the uncertainty of a candidate set of calibration functions. The framework established provides an efficient methodology for chemistry-specific, large-scale investigations of the dynamics and mechanics of epoxy resins. We show preliminary investigations to showcase the model's potential.    

Identifying Accelerated Ageing Pathways for Cross-Linked Polyethylene Pipes using Principal Component Analysis

Cross-linked polyethylene (PEX-a) pipes are widely used for water transport, but their lifetime can be limited by the formation and propagation of cracks in the pipe wall. To disrupt thermo-oxidative degradation pathways and impart long-term stability, stabilizing additives are included in PEX-a pipe formulations. Nevertheless, PEX-a pipes have exhibited cracking and premature failure. We used infrared (IR) microscopy to build a database of high-spatial resolution spectra of cross-sections of pipes subjected to different types of accelerated ageing. The resulting spectra are complex with different correlations between spectral regions and multiple contributions to functional group absorptions. To analyse this convoluted data, we used principal component analysis (PCA) to implement an unsupervised multivariate analytical approach. This allowed us to identify distinct ageing pathways including the detrimental second-order autocatalytic hydrolysis of a key stabilizing additive. The PCA representation of the data allowed us to quantify the kinetics of the hydrolysis reaction and its penetration depth profile into the pipe wall as a function of time and temperature. The results provide important mechanistic information for PEX-a pipe ageing and stabilizing additive formulations.