Jose F. Rodrigues Jr.†, Flavio M. Shimizu‡, Maria Cristina F. de Oliveira†
†Institute of Mathematics and Computer Science, University of São Paulo (USP), CP 668, 13560-970 - São Carlos, SP, Brazil.
‡Brazilian Nanotechnology National Laboratory (LNNano), Brazilian Center for Research in Energy and Materials (CNPEM), CP 6192, 13083-970 - Campinas, SP, Brazil.
Materials Science is increasingly resorting to computational methods to handle the complexity found in the realm of possibilities brought in by applications in all areas of technology. Researchers at both academia and industry are searching for novel high quality materials with designed properties tailored to fit the needs of specific applications. For example, they may seek composite materials possibly resulting from intricate interactions between molecular elements, but with reaction chains that are feasible for deployment in industrial processes.
Following this trend, recent advances in machine learning have been employed to leverage the potential of computers in identifying the patterns governing the behavior of molecules and physical phenomena. The potential benefits have been observed in several domains, from materials prediction to chemical reactivity, passing through quantum calculations. Materials researchers’ long held dreams of discovering novel materials without conducting costly physical experiments might become true in a not so distant future. This would represent a major breakthrough, since decades of intensive research grounded on laboratory experimentation have only scratched the surface of the universe of possible materials that physics can bear.
Nevertheless, a robust scenario in which new materials and reactions can be predicted, rather than being necessarily observed, still depends on finding solutions to numerous problems. First of all, effective Machine Learning relies on substantial amounts of structured high quality data, preferably with labels indicating known facts from which the algorithm will learn the underlying patterns. Then, successful computer algorithms require models that faithfully describe the corresponding real-world system under investigation; at the same time, the complexity of molecular interactions and intrinsic physical properties might easily escalate as the number of molecules and reaction steps increase. Computational issues and open methodological problems also add to the issues that are still to be faced.
Despite the obstacles, it is paramount to pursue strategies to design novel compounds, discover unexpected reactions, in addition to sharpening the interpretation of the data collected from sensors or simulations. The potential social impact of such accomplishments is huge; the findings may point to promising directions for materials research, pave the way for innovation and reshape existing industrial processes.
Here, we resume the special series Shaping the Future of Materials Science with Machine Learning; a new article selection has been compiled reporting recent advances in different areas of Materials Science aiming to guide the reader's experience.
This selection covers discussions on Machine Learning applied to accelerate the design of composite materials and characterize properties. In the paper Mix design factors and strength prediction of metakaolin-based geopolymer; Lahoti et al. employed Machine Learning classifiers to evaluate the mix of design parameters that affect the compressive strength of geopolymers. In another contribution focused on predicting materials properties, viz. High-Throughput Prediction of Finite-Temperature Properties using the Quasi-Harmonic Approximation, Nath et al. proposed a methodology to determine the thermal properties of solid compounds; the authors computed the properties of 130 compounds to demonstrate the method for high-throughput prediction. Still in the domain of thermal properties, Sparks et al. overview data mining and Machine Learning methods for managing information regarding thermoelectric materials; the paper Data mining our way to the next generation of thermoelectrics explains how researchers can gather a comprehensive vision of existing knowledge to develop superior thermoelectric materials.
In the paper An informatics approach to transformation temperatures of NiTi-based shape memory alloys, Xue et al. demonstrated that only three material descriptors related to their chemical bonding and atomic radii suffice to predict the transformation temperatures of shape memory alloys (SMAs); more importantly, the method can accelerate the search for SMAs with desired properties. In Artificial neural network based predictions of cetane number for furanic biofuel additives, Kessler et al. addressed the problem of accelerating the development of alternative fuels, and reported an optimized artificial neural network (ANN) to test a wider variety of fuel candidate types. Li et al., in the paper Feature engineering of machine-learning chemisorption models for catalyst design, considered surface and intrinsic metal properties to engineer numerical models for Machine Learning algorithms; their goal was a rapid screening of transition-metal catalysts.
Based on techniques for predicting materials properties, one can envisage tools targeted at industries concerned with anticipating cracks, leakages, and failures on materials conditioned to friction, temperature or submitted to stressful environments. This type of investigations led to the papers by Thankachan et al., Chou et al., O'Brien et al., and Gould et al., who employ artificial neural networks, support vector machines, classification and regression techniques to find patterns in materials properties in a range of applications. In an interesting approach for crack prevention, Petrich et al., in Crack detection in lithium-ion cells using Machine Learning, apply neural networks to investigate the particle microstructure of lithium-ion electrodes; they use tomographic 3D images to inspect pairs of particles concerning possible breakages. According to Sobie et al., in the paper Simulation-driven machine learning: Bearing fault classification, the accuracy in detecting mechanical faults can benefit from Machine Learning conducted over data acquired from simulations.
Optimizing the entire logistical chain of black top road construction is the aim of the SmartSite project, as discussed in SmartSite: Intelligent and autonomous environments, machinery, and processes to realize smart road construction projects, which employs sensing devices and machine intelligence to increase automation and to monitor processes. In that particular paper, authors focus on intelligent assistance for compactor operators. Another interesting solution that seeks to automate and optimize entire industrial processes is Digitisation of manual composite layup task knowledge using gaming technology; their system captures human actions and their effects on workpieces in manual manufacturing tasks in an industrial setting.
A few reported solutions integrate Machine Learning with techniques of image manipulation for different purposes. The paper 3DSEM++: Adaptive and intelligent 3D SEM surface reconstruction addresses three-dimensional surface reconstruction from two-dimensional Scanning Electron Microscope (SEM) images; other papers handle complex problems on medical imaging to assess the accuracy and efficiency in clinical treatments and diagnosis supported by recent deep learning methodologies, as presented in the following contributions Machine Learning Methods for Histopathological Image Analysis, by Komura and Ishikawa; Role of Big Data and Machine Learning in Diagnostic Decision Support in Radiology, by Syeda-Mahmood; and (Machine-)Learning to analyze in vivo microscopy: Support vector machines, by Wang and Fernandez-Gonzalez.
We expect the compilation presented herein will contribute to foster innovative ideas, illustrate approaches, clarify concepts, and encourage further investigation of Machine Learning applied to the Materials Science research. Advances in this field can accelerate the introduction of innovative processes and applications that might impact the daily lives of many. Nevertheless, despite the impressive advances highlighted, there are still limitations and open issues to be addressed. Scalability remains a challenge, since most applications deal with relatively simple models and small sized systems. Availability and quality of data input to Machine Learning algorithms may also be a critical aspect in some scenarios. Indeed, previous reports of success should not distract researchers into overlooking these and other critical aspects to deploying Machine Learning into systems handling real-world problems. We are not anticipating a scenario in which humans will be replaced by computers in the design of new materials, at least not in a foreseeable future. However, the role played by machine intelligence in empowering humans to handle highly complex problems will continue to grow stronger. Further advances in machine intelligence and optimization of computational models and methodologies will have to accurately and reliably tackle complex application scenarios.
Article Selection
Mix design factors and strength prediction of metakaolin-based geopolymer
M. Lahoti, P. Narang, K. H. Tan, E.-H. Yang
Ceramics International, 2017
High-throughput prediction of finite-temperature properties using the quasi-harmonic approximation
P. Nath, J. J. Plata, D. Usanmaz, R. A. R. A. Orabi, M. Fornari, M. B. Nardelli, C. Toher, S. Curtarolo
Computational Materials Science, 2016
Data mining our way to the next generation of thermoelectrics
T. D. Sparks, M. W. Gaultois, A. Oliynyk, J. Brgoch, B. Meredig
Scripta Materialia, 2016
An informatics approach to transformation temperatures of NiTi-based shape memory alloys
D. Xue, D, Xue, R. Yuan, Y. Zhou, P. V. Balachandran, X. Ding, J. Sun, T. Lookman
Acta Materialia, 2017
Digitisation of manual composite layup task knowledge using gaming technology
V. A. Prabhu, M. Elkington, D. Crowley, A. Tiwari, C. Ward
Composites Part B: Engineering, 2017
An informatics approach to transformation temperatures of NiTi-based shape memory alloys
D. Xue, D, Xue, R. Yuan, Y. Zhou, P. V. Balachandran, X. Ding, J. Sun, T. Lookman
Acta Materialia, 2017
Digitisation of manual composite layup task knowledge using gaming technology
V. A. Prabhu, M. Elkington, D. Crowley, A. Tiwari, C. Ward
Composites Part B: Engineering, 2017
Artificial neural network based predictions of cetane number for furanic biofuel additives
T. Kessler, E. R. Sacia, A. T. Bell, J. H. Mack
Fuel, 2017
Artificial neural network to predict the degraded mechanical properties of metallic materials due to the presence of hydrogen
T. Thankachan, K. S. Prakash, C. D. Pleass, D. Rammasamy, B. Prabakaran, S. Jothi
International Journal of Hydrogen Energy, 2017
Feature engineering of machine-learning chemisorption models for catalyst design
Z. Li, X. Ma, H. Xin
Catalysis Today, 2017
A pattern recognition system based on acoustic signals for fault detection on composite materials
R. J. O'Brien, J. M. Fontana, N. Ponso, L. Molisani
European Journal of Mechanics - A/Solids, 2017
SmartSite: Intelligent and autonomous environments, machinery, and processes to realize smart road construction projects
R. Kuenzel, J. Teizer, M. Mueller, A. Blickle
Automation in Construction,2016
From machine learning to deep learning: progress in machine intelligence for rational drug discovery
L. Zhang, J. Tan, D. Han, H. Zhu
Drug Discovery Today, 2017
3DSEM++: Adaptive and intelligent 3D SEM surface reconstruction
A. P. Tafti, J. D. Holz, A. Baghaie
Micron, 2016
Role of Big Data and Machine Learning in Diagnostic Decision Support in Radiology
T. Syeda-Mahmood
Journal of the American College of Radiology, 2018
Crack detection in lithium-ion cells using machine learning
L. Petrich, D. Westhoff, J. Fein, D. P. Finegan, S. R. Daemi, P. R. Shearing. V. Schmidt
Computational Materials Science, 2017
Classification of failure mode and prediction of shear strength for reinforced concrete beam-column joints using machine learning techniques
S. Mangalathu, J.-S. Jeon
Engineering Structures, 160 (2018)
(Machine-)Learning to analyze in vivo microscopy: Support vector machines
M. F. Z. Wang, R. Fernandez-Gonzalez
Biochimica et Biophysica Acta (BBA) - Proteins and Proteomics
Machine learning in concrete strength simulations: Multi-nation data analytics
J.-S. Chou, C.-F. Tsai, A.-D. Pham, Y.-H. Lu
Construction and Building Materials, 2014
Thermal response construction in randomly packed solids with graph theoretic support vector regression
D. W. Gould, H. Bindra, S. Das
International Journal of Heat and Mass Transfer, 2017
New prediction models for concrete ultimate strength under true-triaxial stress states: An evolutionary approach
S. K. Babanajad, A. H. Gandomi, A. H. Alavi
Advances in Engineering Software, 2017
Simulation-driven machine learning: Bearing fault classification
C. Sobie, C. Freitas, M. Nicolai
Mechanical Systems and Signal Processing, 2018
Bayesian optimization for efficient determination of metal oxide grain boundary structures
S. Kikuchi, H. Oda, S. Kiyohara, T. Mizoguchi
Physica B: Condensed Matter, 2018
A framework for data-driven analysis of materials under uncertainty: Countering the curse of dimensionality
M. A. Bessa, R. Bostanabad, Z. Liu, A. Hu, D. W. Apley, C. Brinson, W. Chen, W. K. Liu
Computer Methods in Applied Mechanics and Engineering, 2017
Comparative analysis of data mining and response surface methodology predictive models for enzymatic hydrolysis of pretreated olive tree biomass
F. Charte, I. Romero, M. D. Pérez-Godoy, A. J. Rivera, E. Castro
Computers and Chemical Engineering, 2017
Data driven modeling of plastic deformation
D. Versino, A. Tonda, C. A. Bronkhorst
Computer Methods in Applied Mechanics and Engineering, 2017
Differentiation of Crataegus spp. guided by nuclear magnetic resonance spectrometry with chemometric analyses
J. A. Lund, P. N. Brown, P. R. Shipley
Phytochemistry, 2017