Machine Learning & Robotics in New Material Discovery: Innovations, Start-Ups, Applications

​

Tailor-making materials for a given application is one of the most desired, yet challenging, technological advancements of our century. We need these materials to reach the global sustainability goals of our society, including climate action and affordable clean energy. The success in generating large quantities, high-quality data on materials in the past decade makes the field ready for an abrupt growth toward this aim by applying the tools from the field of artificial intelligence. To enable this, however, we need to develop material-specific machine learning approaches and methodologies. In my talk, I will discuss how we are approaching this challenge by discussing a few success stories from the field of nanoporous materials for energy applications, including quantifying the novelty of new materials, learning from failures, and multi-scale design from atoms to chemical plants.

​

As materials and manufacturing processes get more and more complicated also R&D processes become more complex. Traditional R&D methods are often too inefficient to harness the full potential of these new materials and manufacturing processes. Machine learning based R&D software is faster, more efficient and offers many other benefits. However, many ML tools are built from data scientists for data scientists, hence, are complicated to use and require user expertise. In my talk I will show you how easy to use tools can help engineers and researchers to reduce R&D time by up to 95% and mitigate supply chain risks without the need of ML or programming knowledge.

Transforming multi-dimensional, unstructured, and dispersed materials data into AI/ML driven results for R&D, supply chain and manufacturing is a challenge. Doing so rapidly, cost-effectively, and sustainably on a collaborative organizational platform, is an even bigger challenge. Like the “Rolodex-to-CRM” evolution in marketing/sales before it, the Materials Informatics Platform (MIP) is the next organizational platform evolution for materials/products. We will showcase via solid-state batteries and perovskite solar cells from the www.perovskitedatabase.com, although domain agnostic and proven on hydrogen cells, building materials, polymers, 3D printing, alloys, coating, packaging, healthcare products and more.

​

Machine learning (ML) and Artificial Intelligence (AI) have made significant advances in chemical design. Autoencoders and Convoluted Neural Networks have allowed the abstraction of chemical structures into defined latent space that can be optimized or correlated to complex chemical properties. This has led to the assumption that ML will enable rapid design of new chemical compounds. Unfortunately, ML models often require hundreds of thousands of well-defined data points for training, which is often difficult to get experimentally, or impossible for novel materials. Due to the lack of experimental data, ML models are often trained on data from quantum chemistry simulations such as Density Functional Theory. ML models trained on quantum chemistry simulations inherit the issues embedded into these quantum chemistry theories without careful consideration of the bias that exists within quantum chemistry methods, and thus only accelerates our convergence to the wrong answer. This requires scientists to develop accurate quantum chemistry methods to generate accurate synthetic data to support ML algorithms. Here we present a novel quantum computing inspired method – the iterative Qubit Coupled Cluster method - that shows strong correlation between functional group changes and observed phosphorescent emission in Organic Light Emitting Diodes. We will discuss how these methods can be used to generate accurate and consistent databases for ML training.

Kyulux has been developing emissive small molecules for OLED devices for the past few years. In Particular, we have been trying to build a comprehensive web-based platform that can cover a wide range of key elements for material informatics, including experimental/computational data collection, automated data pre-processing, machine learning using those data, and data visualization. A few practical challenges we have faced will be discussed.

Despite the intriguing properties and widespread applications of semiconductor nanomaterials in energy technologies, their discovery, synthesis, and manufacturing are still based on Edisonian trial-and-error based techniques. Existing materials development strategies using resource-intensive batch reactors with irreproducible and uncontrollable heat/mass transport rates very often fail to overcome the demands of the vast synthesis and processing universe of energy materials, resulting in a slow and expensive discovery and development timeframe (8-10 years). Recent advances in lab automation and machine learning (ML)-guided modeling/decision-making strategies provide an exciting opportunity to reshape the discovery and manufacturing of emerging solution-processed energy materials.1 In this talk, I will present an end-to-end ‘self-driving lab’ for autonomous discovery, development, and on-demand manufacturing of energy materials through convergence of flow chemistry, robotics, and in-situ material characterization with ML.2 I will discuss how modularization of different stages of materials synthesis and processing in tandem with an ensemble neural network modeling and decision-making under uncertainty can enable a resource-efficient navigation through an experimentally accessible high dimensional space. An application of the self-driving lab for the autonomous synthesis of metal halide perovskite quantum dots will be presented

Many important mechanical properties can only be measured using physical experiments, which means that design for such properties happens through a slow and expensive iterative cycle. Here, we describe our efforts to reinvent this paradigm using autonomous experimentation. In particular, we combine robots to perform experiments in an automated and reliable fashion with machine learning to select each subsequent design. Benchmarking has revealed that this process converges on high performing structures nearly 60 times faster than conventional grid searching. We discuss how autonomous experimentation accelerates the design process and allow us to identify tough and resilient structures for numerous applications ranging from personal protective equipment to crumple zones on cars.

Chemical/material formulations are characterized by large numbers of components but also a diversity of interacting and competing forces that determine system properties, making them difficult to model. Most formulated products are complex, but even after decades of development they are still designed using a combination of experience and heuristics. Hierarchical machine learning (HML) was developed as a tool for designing these systems based on the small datasets that are common in research and development.

HML is both an algorithm for designing and a process for understanding complex chemical/material systems. The goal in machine learning is to generate a response surface that accurately relates the input variables with the output responses of a system. HML generates a second response surface that is parameterized by latent variables that represent the underlying forces and interactions. These latent variables are obtained from theoretical or empirical models or surrogate physical measurements and represent conceptual understanding of the system. However, in contrast with heuristic approaches that are often reductionist, HML retains the full complexity of the forces and interactions that govern the system properties. It provides a path to both optimized design based on input parameters and their constraints as well as conceptual understanding of how these systems work, thus serving as a tool for applied research and development.

Ansatz AI has applied HML to a diversity of industrial technologies with corporate clients, ranging from molecular engineering of polymeric elastomers, lubricants, and dielectrics to liquid formulations of paints, coatings and personal care products. In these technologies, the goals have ranged from increasing performance to minimizing costs to shifting to renewable feedstocks. This seminar will explain the HML approach to modeling complex chemical/material formulations and highlight some of the solutions that it has provided.

​

Metal-organic frameworks (MOFs) are a versatile class of nanoporous materials synthesized in a “building-block” approach from inorganic nodes and organic linkers. By selecting appropriate building blocks, the structural and chemical properties of the resulting materials can be finely tuned, and this makes MOFs promising materials for applications such as gas storage, chemical separations, sensing, drug delivery, and catalysis. This talk will focus on efforts to design or screen MOFs for storing natural gas (which is primarily methane) for applications such as transportation. Because of the predictability of MOF synthetic routes and the nearly infinite number of possible structures, molecular modeling is an attractive tool for screening new MOFs before they are synthesized. Modeling can also provide insight into the molecular-level details that lead to observed macroscopic properties. This talk will illustrate how a combined modeling and experimental approach can be used to discover, develop, and ultimately design new MOFs for gas storage and other applications.

A material design strategy is proposed combining machine learning models with optimization algorithms to guide the design of novel High-Entropy Alloys (HEAs) optimized for High-Temperature strength and Density. The data used to train machine learning model was generated using Ab-initio Molecular Dynamics, and contained composition of the alloys, their mechanical properties, and relevant chemical descriptors previously identified in literature. This Rapid Alloy Design strategy can be used to optimize the properties of other multi-component alloys such as Bulk Metallic Glasses and Nickel Superalloys.

A transfer learning approach for the automatic segmentation of microscopy data is presented. Many encoder architectures, including VGG, Inception, ResNet, and others, were trained on ~100,000 microscopy images from 54 material classes to create pre-trained models that learn representations that are more relevant to downstream microscopy analysis tasks than models pre-trained on natural images. The pre-trained encoders were embedded into segmentation architectures including U-Net and DeepLabV3+ to evaluate the performance of models pre-trained on a large microscopy dataset. Each encoder/decoder pair was evaluated on several benchmark datasets. Our testing shows that models pre-trained on a large microscopy dataset generalize better to out-of-distribution data (micrographs taken under different imaging or sample conditions) and are more accurate when training data is limited than models pre-trained on ImageNet. The application of this technique to segment and quantify microscopy features from Ni-superalloys and environmental barrier coatings is demonstrated.

Most commercial refractory alloys were designed with high temperature strength and manufacturability prioritized over oxidation resistance. This drives the need for complex and expensive coatings in aggressive service conditions. By lifting classic composition constraints through a high entropy alloying approach, it is possible to achieve improved balance-of-properties in refractory metals. Tailoring properties individually, as required for a specific application or as input for design trades, is also enabled. Here, we review recent work combining high throughput experimental screening, machine learning, and multi-objective optimization to explore a wide refractory alloy composition space. We demonstrate a materials informatics alloy selection process for extreme service conditions where oxidation resistance is prioritized alongside mechanical properties and manufacturability. The general methods presented here can be applied to other applications and highlight the benefits of a materials informatics approach to alloy design.

The advancement of transformative technologies for building a sustainable future requires significant innovations in the design, development, and manufacturing of advanced materials and chemicals. To meet global sustainability needs faster, we must transition to the data-centric approaches of materials informatics. Materials informatics leverages the synergies between materials science, data science, and artificial intelligence and provides the foundations of a paradigm shift in materials discovery and development.
By increasing the agility of research and accelerating the development cycles, materials informatics enables transformative discoveries to reduce emissions and resource intensity, improve energy efficiency and circularity, expand the market share of sustainable products, and increase responsible raw material sourcing.

​

Accelerated materials discovery has long been a stated goal of our research community, and organized efforts towards this goal are numerous and well-supported. Furthermore, simulation and machine learning have rapidly become popular methods of rationalizing and predicting material properties. However, the full-stack process of generating new material candidates, predicting their properties, synthesizing them, and identifying the resulting structure and function has not yet reached a widespread inflection point in its efficiency. In this talk, I describe Toyota Research Institute’s (TRI) end-to-end process, enhanced by AI and simulation, in which we have discovered previously unobserved inorganic crystal structures. Synthesis and characterization remain rate-limiting in our process, but preliminary research shows promise in addressing these bottlenecks to realize an accelerated full-stack materials discovery process.

There is a growing demand for novel materials to resolve global problems involving sustainability, health, and global warming. These pressing issues result in an unprecedented urge and pressure to be faster and more cost-efficient in running material R&D programs. In this presentation, I show how Kebotix accelerates innovation by shifting from a traditional disconnected and serendipitous R&D approach to a tightly integrated closed-loop design-make-test process.

The engine powering our process is the Kebotix self-driving lab, where we combine materials modeling, artificial intelligence, data management, and lab automation into an integrated platform. The software prioritizes tested ideas, orchestrates each step in the workflow, and collects data on the fly. Our digital technologies increase the operational efficiency of R&D programs by up to an order of magnitude, as we demonstrate in various success stories, ranging from accelerated materials formulation design to novel material discovery for optoelectronics applications such as smart windows, OLEDs, or Fluorescent Biomarkers.

We have entered a paradigm-changing era in the way chemists innovate. Many fields, such as automotive engineering and particle physics, have relied on accurate simulation before experimentation for many years. In recent years, chemistry has entered a new phase of materials design powered by the synergy of physics-based tools, informatics, and augmented intelligence (AI) capabilities that work together synergistically; accelerating materials development from atomistic to coarse-grained simulation.

In this talk, critical case studies illustrating the latest physics-based simulation technology combined with advanced machine learning for the design and optimization of next-generation materials will be discussed. Underlying these success stories are a combination of materials science-specific descriptors, interpretable machine learning models, and most importantly, the critical role of physics-based simulations coupled with generative methods to automatically design new materials solutions. Finally, we will also introduce an enterprise informatics platform focused on chemical discovery, enabling multidisciplinary teams to amplify their development cycle with collaboration on a global scale.

​

​

Accelerating the discovery and commercialization of novel materials is necessary for maintaining economic competitiveness and timely addressing many societal issues (e.g., sustainable manufacturing and clean energy). For several decades now, simulations have complemented empirical science for such acceleration, culminating in several successful industrial applications of this approach, termed integrated computational materials engineering (ICME). In 2011, the Materials Genome Initiative (MGI) sought to apply this idea at scale across all materials industries, including a third “digital data” pillar. Materials informatics is the practical manifestation of “digital data” methods to materials science problems, including: (1) the collection, generation, and distribution of materials data, (2) the use of that data to train machine learning models for predicting process-structure-property relationships, and (3) the design of experiments using artificial intelligence (AI) algorithms based on those models.
Citrine Informatics is a software company building a scalable, enterprise-level materials informatics platform for data-driven materials and chemicals development. The Citrine Platform combines smart materials data infrastructure and AI, which accelerates development of cutting-edge materials, facilitates product portfolio optimization, and codifies research IP, enabling its reuse and preventing its loss. Citrine's customers include Panasonic, Michelin, LANXESS, and others in the materials, chemicals, and product manufacturing industries.
In this talk, the concepts around materials informatics will be introduced, Citrine’s software will be described, and several case studies demonstrating the value of materials informatics will be discussed.

​

Technology progresses only as fast as the development of new, advanced materials. Modern alloys, plastics, composites, and electronic materials have made possible the impossible in countless applications, but the discovery of these materials has been way too slow for way too long. However, over the last decade there has been a paradigm shift away from trial-and-error or rudimentary design rules towards the adoption of materials informatics to greatly accelerate the discovery and development timeline.
This talk will discuss the premise behind materials informatics and will showcase a few key examples of ways that materials informatics is transforming the field with case studies in new superhard materials, 3D printable alloys, bulk metallic glasses, new LED materials and more.
I will then talk about some of the remaining challenges in this field that need to be overcome in order to unlock materials informatics to its full potential. Specifically, I will describe limitations of modern algorithms, materials representations, the difference between screening and discovery, and out-of-domain machine learning tasks.

Autonomous experimentation has had a significant impact on how many large-scale experimental facilities operate, however; the concept is often linked to mystery and confusion. In this talk, I will take a very practical look at the subject. I will introduce Gaussian-process-driven autonomous experimentation on a very high level, followed by the presentation of a few examples from across large-scale experimental facilities. This talk is also meant as guidance for everyone in the audience who wants to use autonomous experimentation for their research and will present some readily-available tools.