Artificial Intelligence for Materials Science

Considerable effort has been devoted to the research of Materials Genome Initiative (MGI), which requires the optimization of multiple interrelated properties, which are critical for developing advanced materials. As big data involved in the simulations and the experiment, the understanding of the MGI remains challenging. The machine learning methods, which have been adopted in the MGI, developed with big data and artificial intelligence. This chapter provides a brief overview of the machine learning methods adopted in the materials studies.

Metals and alloys have been acting an irreplaceable role in human beings’s civilization progress. Since high-entropy alloys (HEAs) were proposed in 2004, tremendous progresses and profound developments have been made in both the fundamental investigations and engineering applications. Unlike the conventional metallic alloys that typically only consist of one or two principal elements, HEA is composed of multi-principal elements in equimolar or near-equimolar compositions. The huge compositional space provides a great opportunity to discover HEAs with desired properties, such as outstanding structural stability, a fascinating balance between strength and ductility, unique wear resistance, and superior thermal stability. Meanwhile, this huge compositional space is also challenging for conventional material discovery procedure. Machine learning (ML) approaches, which are computationally inexpensive, highly efficient, and easily transferable, have been employed to accelerate HEA development. This chapter will give an overview of HEAs (fundamentals, preparations, and properties) and introduce recent progress in ML-assisted design of HEAs (microstructure and property predictions).

In this chapter, computational inverse techniques based on trumpet neural networks (TrumpetNets) and tube neural networks (TubeNets) are introduced for identification of material constants. An idealized case of laminated composites is considered that may have a large number of material constants need to be determined, including Young’s modulus, Poisson’s ratio, and shear modulus for different plies in the laminate. The TrumpetNets (or TubeNets) consists of both forward and inverse neural networks (NNs) with data flows in two ways. The responses in terms of displacements and/or strains on the surface (so that measurable) of composite laminates are used as the inputs for the inverse NN, and the outputs of the inverse NN are these material constants. Numerical method, such as the finite element method (FEM), is used as the forward solver to calculate the displacement and strain responses of the composite laminated structures with a given set of material constants and subject to given loading boundary conditions. The forward NN is trained using the datasets generated using FEM. The trained forward NN is then used to train the inverse NN. The superiority of the TrumpetNet and TubeNets is the high effectiveness for both forward and inverse problems due to the two-way architecture. In addition, the direct-weight-inversion (DWI) method may also be used, for some cases, to compute directly all the weights and biases for each and every layer in the inverse NN (Liu, Int J Comput Methods 16:1950045, 2019; Liu et al., Int J Comput Methods 18:2050030, 2021). In this case the expensive training of the inverse NN can be avoided, and hence the efficiency can be improved further.

Recently, the machine learning (ML) atomic force field has emerged as a powerful atomic simulation approach because of its high accuracy, low computational cost, and transferability. In this mini review, we first summarize the disadvantages of traditional force field and the unique advantages of ML-based force field for molecular dynamics simulations. Then the basic workflow to develop the ML atomic force field is discussed in each step. Furthermore, taking carbon material as a typical example, the various applications of ML-based force fields for studying the carbon allotropic materials are reviewed. Finally, the perspectives are discussed and future directions for studying atomic force field by ML are given.

Genetic algorithm, a very simple but very powerful stochastic global optimizer, has been applied to many fields in search and optimization. This capture study aims to provide overall information for a genetic algorithms user to choose the most appropriate scheme for his or her specific application problem for his or her specific application problem. In this section, we mainly address three well-known genetic algorithms, namely, artificial bee colony, particle swarm optimization, and differential evolution. The evolution mechanism, current research status, and applications of different genetic algorithm have been investigated in detail for the users to choose the most appropriate strategy.

Optimized electronic and thermal transport properties are the key requirements for the discovery of efficient thermoelectric materials. Owing to the complex interdependence, simultaneous optimization of these properties is a non-trivial and challenging task, especially if one wants to explore the large available search space of materials. With the advent of statistical high-throughput and machine learning based approaches, several of these challenges for thermoelectrics have been addressed. The goal of this chapter is to highlight these data-assisted efforts towards accelerated development of high-performance thermoelectric materials. We will discuss the contribution of curated databases for high-throughput screening of desired electronic and thermal transport properties. The utilization of these databases will also be described for development of prediction models of transport properties, which has accelerated the discovery of highly efficient thermoelectric materials. Details of commonly used strategies and methods to select a relevant descriptor set for developing the prediction models will be covered. A new approach for selecting descriptors by analyzing the high-throughput property map will be explained. The potential of machine learning methods in relating the unrelated properties will be illustrated by establishing a connection between otherwise independent electronic and thermal transport properties. Further, for designing the highly transferable models for a single target property of interest, we will also cover localized regression based algorithmic development.

Tuning thermal transport by nanostructures has garnered increasing attentions as thermal materials with either high or low thermal conductivities are of great use in a wide range of applications like thermal management, thermal barriers, and thermoelectrics. Due to the superhigh degree of freedoms in terms of atom types and structural configurations, traditional searching algorithm may be powerless to find the optimal nanostructures with limited time and computation expenses, and thus the big-data-driven materials informatics (MI), as the fourth paradigm, has emerged and become prevalent. In this chapter, beginning with the brief introduction of the MI algorithms, we emphasize on the progress on MI-based thermal nanostructure designs, ranging from heat conduction through Si/Ge and GaAs/AlAs superlattices, graphene nanoribbons, to thermal emission for radiative cooling, ultranarrow emission, thermophotovoltaic system, and thermal camouflage. The remaining challenges and opportunities in this field are outlined and prospected.

In recent years, lead-halide perovskite (LHP) have made tremendous progress in photovoltaic and optoelectronic fields. However, stability and toxicity still are obstacles for commercial application. These challenges have motivated significant efforts to search nontoxic and stable alternatives which could achieve comparable high performance with low-cost and facile fabrication methods. With continuing increasing computation powers, first-principles modeling combining with machine learning (ML) has made significant advances in the discovery of new perovskite materials. This provides invaluable insights into the physical origin for the high performance of perovskites in photovoltaic field, thereby facilitating the discovery of good absorber perovskite materials. This chapter aims to give a brief review of ML-guided design and discovery of perovskite materials for photovoltaics. Specifically, this chapter will introduce well-established ML models widely used in perovskite-related studies from both the construction of data and material representation aspects. The approaches of data sets will be discussed including the high-throughput (HT) computations and experimentations. The material representation will cover descriptors and feature engineering of perovskites in photovoltaic field. Then, we will give a general introduction of recent progress for ML models applications in perovskite solar cells. Conclusion and outlook will be given in the end.