Session C60: AI Materials Design and Discovery II

Live

We build random forests models to predict mechanical properties of a compound, using only its chemical formula as input. The model is trained with over 10,000 target compounds and 60 features based on stoichiometric attributes, elemental properties, orbital occupations, and ionic bonding levels. The model can achieve Pearson correlation coefficients > 0.9 for bulk and shear modulus regressions. Using the model, we construct triangular graphs for B-C-N compounds to map out their bulk and shear moduli, as well as hardness values. The graphs indicate that a 1:1 B-N ratio can lead to various superhard compositions. We also apply the machine learning models to search new superhard compounds, and validate the results using evolutionary structure prediction and density functional theory. We discover several dynamically stable phases of B-C-N compounds with hardness values > 40GPa, which are potentially new superhard materials that could be synthesized by low-temperature plasma methods.   

Live

When designing a multi-layer core-shell nanoparticle for a desired optical response, it is necessary to understand what key structural parameters are at play. Here we utilize a neural network to train, and subsequently compute Mie optical responses for multi-layer nanoparticles, consisting of amorphous silicon as the core and silicon dioxide as the coating. Once trained, the neural network can efficiently simulate optical scattering responses faster than traditional transfer matrix method. Finally, our neural network is used to solve the inverse design problem in order to develop an understanding of the structural parameters (diameter, layer thickness and dielectric) that affect the quality factors of the nanoparticle optical response. Furthermore, we provide insights on how different loss functions used in the search algorithm can lead to profound differences in the optimization process, while still providing accurate optical spectra for our core-shell nanoparticles.   

Live

Polymers, due to advantages such as low-cost processing, chemical stability, low density and tuneable design, have emerged as a powerhouse class of materials. However, precisely because the design space is so large, traditional approaches (be it experiment or pure simulation) for identifying application-specific polymers are often infeasible: they simply take too long. To accelerate the search, we need a radically different approach, the most promising of which are driven by artificial intelligence, AI, and therefore offer ultrafast predictions.

Here, we present a novel AI platform for the generative design of polymers and use it to discover promising dielectric materials. The key insight is that the distribution of subtle chemical differences between high- and low-performing (as measured by property objectives) polymers can be learned and sampled to generate hypothetical, high-performing materials. Our AI finds tens of thousands of dielectric polymers which meet extreme objectives. Density functional theory simulations of bandgap and electron injection barrier confirm that, out of a small subset, 50% of these hypothetical polymers do indeed match the objectives. Finally, we uncover design rules from the AI and present them as potential structure-property relationships.   

Live

The gas-phase molecular dipole moment is a central quantity in chemistry. It is essential in predicting molecular infrared and sum-frequency-generation spectra, as well as in describing long-range interactions. Here we fit a machine learning model on an accurate quantum chemical reference dipole set. We represent the dipole with a physically-inspired machine learning model that captures the two distinct physical effects contributing to molecular polarization: Local atomic polarization is captured within the symmetry-adapted Gaussian process regression (SA-GPR) framework, while long-range movement of charge is captured by assigning a scalar charge to each atom. Not only does the model achieve state-of-the-art interpolation and extrapolation performance on the standard QM9 reference set, it also gives useful insights into the physics of polarization and charge transfer for a variety of challenging test examples. The results show how transparency and physical interpretability can aid not only the understanding of a machine learning model, but allow it to achieve higher accuracy as well. Extensions to the condensed phase, within the context of the modern theory of polarization, are discussed.   

Live

An essential component of materials research is the use of simulations based on density functional theory (DFT), which imposes severe limitations on the size of the system under study. A promising development in recent years is the use of machine learning (ML) methodologies to train surrogate models with DFT data to predict quantum-accurate results for larger systems. Many successful ML models have been created to predict higher-level DFT results such as the total potential energy and atomic forces, and initial steps have been taken to create machine-learning based ML methodologies that can predict fundamental DFT outputs such as the charge density, wave functions and corresponding energy levels. Here, we explore the applicability of this latter methodology using deep learning neural networks to learn and predict the electronic structure of carbon, for a large variety of allotropes [1], and its extension to hydrocarbon molecules and polymers. Further improvements to the speed, accuracy and versatility of this DFT-emulation methodology will also be presented.

REFERENCES:
[1] B.G. del Rio, C. Kuenneth, H. Tran, and R. Ramprasad, J. Phys. Chem. A, accepted (2020)   

Live

We create a new dataset composed of over 1500 non-fullerene organic solar cells (NF-OSCs) by curating experimental data from recently published literature. The dataset includes performance metrics such as power conversion efficiencies (PCEs), short-circuit currents, open-circuit voltages, and fill factors of the NF-OSCs, together with chemical structures of donor/acceptor pairs and device fabrication conditions, which have been reported to have a significant influence on the device performance. Additionally, we conduct quantum chemical calculations of donor/acceptor molecules present in the dataset to obtain their electrochemical properties. We construct several features reflecting chemical structures, electrochemical properties, and device fabrication conditions, which are subsequently fed into a kernel ridge regression model to predict the PCEs of the NF-OSCs. The prediction results indicate that the structural feature alone is insufficient for reliable prediction, while concatenating structural and electronic features significantly improves the prediction performance. We also examine the prediction performance for out-of-sample NF-OSCs containing Y6-type acceptors, highlighting reasonable efficacy of the concatenated features.   

Live

We have developed a Machine Learning (ML) framework for modeling the absorption spectra of azobenzene molecules—an important class of light-absorbing compounds with many current and potential applications. The ML models utilize predictors based on the structure and composition of each azobenzene molecule. Due to the relatively small size of the dataset (less than 500 molecule-spectrum pairs), dimensionality reduction of the original predictors is an important feature extraction step. With the reduced set of predictors, we trained separate regression models to predict the absorption at different wavelengths in the UV – visible light range. These models are able to accurately predict the absorption at fixed wavelengths, as well as the position and intensity of the maximum absorption. These predictions can be used to rapidly screen thousands of candidate molecules for a variety of potential applications, reducing the need for time-consuming and expensive experiments or quantum chemistry computations.   

Live

We present a machine learned model for predicting the volume of a homomolecular crystal from the single molecule structure. The model is based on two descriptors: the volume enclosed by the packing-accessible surface and molecular topological fragments. To calculate the volume enclosed by the molecular surface we have developed a new "projected marching cubes" algorithm. The packing-accessible surface is then calculated using an optimized probe radius. The molecular topological fragments are used to construct a representation that captures the bonding environments of the atoms in the molecule. Feature selection is used to determine which fragments to include in the model. The accuracy and robustness of the model may be attributed to including both geometric and chemical features. The volume enclosed by the packing-accessible surface accounts for the presence of voids and sterically hindered regions as well as for the effect of conformational changes. The molecular topological fragments account for the effect of intermolecular interactions on the packing density. The model is trained on a dataset of structures extracted from the Cambridge Structural Database. Excellent performance is demonstrated for three validation sets of unseen data.   

Live

Predicting the outcome of a catalytic reaction is of relevance to high-throughput screening of chemical reactions for industrial applications. High-throughput screening can significantly reduce the number of experiments needed to be performed in a huge search space, which saves time, effort and expense. In this talk we show that machine learning can be used to accurately predict the outcomes of catalytic reactions on the surface of oxygen-covered and bare gold in a database. Our machine learning models exploit a chemical space representation of the molecules in the database. Studying the catalytic reactions in this chemical space may provide insights into their behavior. Furthermore, our approach provides a framework for performing high-throughput screening of chemical reactions, as well as venues for pursuing the inverse design of industrially relevant molecules. Our machine learning framework complements chemical intuition in predicting the outcome of several types of chemical reactions. In some cases, machine learning makes correct predictions where chemical intuition fails. We achieve up to 93% prediction accuracy for a small data set of less than two hundred reactions.   

Live

Graphene is a 2D material which shows many promising applications in diverse technologies such as sensors, field effect transistors, and solar cells. Graphene can be formed into quasi-1D nanoribbons which show further device application. The optoelectronic properties of graphene nanoribbons are readily manipulated through multiple approaches. We demonstrate engineering of the optical response of graphene nanoribbons using density functional theory to compute bandgaps and dielectric functions and machine learning. Structure, width, strain, electronic doping, edge functionalization, and point-substitutions are included as methods to control the optoelectronic properties of nanoribbons. Nanoribbon structures are procedurally generated and evaluated for electronic bandgap and dielectric function. Machine learning is used to identify trends between nanoribbon structure and optical properties and predict new structures with desired optical response.   

Not Participating

Over ten orders of magnitude separate the microscopic processes and macroscopic equilibration time for spin ices such as Dy₂Ti₂O₇. Understanding the slow, stochastic, dynamics of these systems is a problem on which machine learning may be able to lend some new insight. To this end, we have generated datasets consisting of kinetic Monte Carlo simulations of both two-dimensional artificial spin ice and three-dimensional pyrochlore spin ice.

In the interest of increasing the scale of simulations, we've implemented a convolutional neural network which contains no dense layers. This means that the model is agnostic to input-size and thus highly scalable. The kernel size of our model is also minimal, only permitting nearest-neighbor interaction, yet the model reproduces the principle dynamics quite well. This suggests that, to some extent, the rules governing the dynamics of spin ice are primarily local and that the scalability of the 'ice rules' may be a viable route to solving the spin ice problem. This also demonstrates that a deterministic neural network is capable of learning the stochastic time-series of a complex physical system, a general problem.   

Not Participating

Most computational materials databases are currently limited to properties calculated at zero temperature, but the inclusion of temperature can vastly change the energetic landscape, influencing our predictions of what compounds we believe to be stable and synthesizable. However, the high cost of computing temperature-dependent properties prohibits large high-throughput studies from being performed. Here, we train a simple machine-learning model to efficiently predict the vibrational entropy and free energy of materials from composition alone. Other previous studies include similar models trained on small datasets of hundreds of compounds at only a single temperature, but our model was trained on a set of thousands of compounds and achieves better accuracies over a broad range of compositions, temperatures, and structural complexities. The accuracy and low computational cost of this approach make it possible to generate temperature-dependent phase diagrams for numerous systems, providing insight into the effect of temperature on stability.