Learning and intelligent optimization for material design innovation mosavi

Page 1

Learning and Intelligent Optimization for Material Design Innovation Amir Mosavi1,2(&) and Timon Rabczuk1(&) 1

Institute of Structural Mechanics, Bauhaus-Universitat Weimar, Marienstr.15, 99423 Weimar, Germany {amir.mosavi,timon.rabczuk}@uni-weimar.de 2 Department of Computer and Information Science, Norwegian University of Science and Technology, Sem Saelandsvei 9, 7491 Trondheim, Norway

Abstract. Learning and intelligent optimization (LION) techniques enable problem-specific solvers with vast potential applications in industry and business. This paper explores such potentials for material design innovation and presents a review of the state of the art and a proposal of a method to use LION in this context. The research on material design innovation is crucial for the long-lasting success of any technological sector and industry and it is a rapidly evolving field of challenges and opportunities aiming at development and application of multi-scale methods to simulate, predict and select innovative materials with high accuracy. The LION way is proposed as an adaptive solver toolbox for the virtual optimal design and simulation of innovative materials to model the fundamental properties and behavior of a wide range of multi-scale materials design problems. Keywords: Machine learning

Optimization Material design

1 Introduction Materials design is crucial for the long-lasting success of any technological sector, and yet every technology is founded upon a particular materials design set. This is why the pressure on development of new high-performance materials for use as high-tech structural and functional components has become greater than ever. Although the demand for materials is endlessly growing, experimental materials design is attached to high costs and time-consuming procedures of synthesis. Consequently simulation technologies have become completely essential for material design innovation [1]. Naturally the research community highly supports the advancement of simulation technologies as it represents a massive platform for further development of scientific methods and techniques. Yet computational material design innovation is a new paradigm in which the usual route of materials selection is enhanced by concurrent materials design simulations and computational applications [19]. Designing new materials is a multi-dimensional problem where multiple criteria of design need to be satisfied. Consequently material design innovation would require advanced multiobjective optimization (MOO) [13] and decision-support tools [12]. In addition the performance and behavior of new materials must be predicted in © Springer International Publishing AG 2017 R. Battiti et al. (Eds.): LION 2017, LNCS 10556, pp. 358–363, 2017. https://doi.org/10.1007/978-3-319-69404-7_31


Learning and Intelligent Optimization for Material Design Innovation

359

different design scenarios and conditions [2]. In fact predictive analytics and MOO algorithms are the essential computation tools to tailor the atomic-scale structures, chemical compositions and microstructures of materials for desired mechanical properties such as high-strength, high-toughness, high thermal and ionic conductivity, high irradiation and corrosion resistance [7]. Via manipulating the atomic-scale dislocation, phase transformation, diffusion, and soft vibrational modes the material behavior on plasticity, fracture, thermal, and mass transport at the macroscopic level can be predicted and optimized accurately [17]. Therefore the framework of a predictive simulation-based optimization of advanced materials, which yet to be realized, represents a central challenge within material simulation technology [9]. Consequently material design innovation is facing the ever-growing need to provide a computational toolbox that allows the development of tailor-made molecules and materials through the optimization of materials behavior [10]. The goal of such toolbox is to provide insight over the property of materials associated with their design, synthesis, processing, characterization, and utilization [19].

2 Computational Materials Design Innovation Computational materials design innovation aims at development and application of multiscale methods to simulate advanced materials with high accuracy [17]. A key to meet the ever-ongoing demand on increasing performance, quality, specialization, and price reduction of materials is the availability of simulation tools which are accurate enough to predict and optimize novel materials on a low computation cost [6]. A major challenge however would be the hierarchical nature inherent to all materials. Accordingly to understand a material property on a given length and time scale it is crucial to optimize and predict the mechanisms on shorter length and time scales all the way down to the most fundamental mechanisms describing the chemical bond. Consequently the materials systems are to be simultaneously studied under consideration of underlying nano-structures and Mesomanufacturing Scales. Such design process is highly nonlinear and requires an interactive MOO toolset [12]. 2.1

Interdisciplinary Research and Research Gap

Structure calculations of materials [20], systematic storage of the information in database repositories [8], materials characterization and selection [18], and gaining new physical and environmental insights [9] account for big data technologies. In addition making decision for the optimal materials design needs MOO tools as well as an efďŹ cient decision-support system for post-processing [21]. This is considered as a design optimization process of the microstructure of materials with respect to desired properties and Mesoscale functionalities. Such process requires a smart agent which learns from dataset and makes optimal decisions. The solution of this inverse problem with the support of the virtual test laboratories and knowledge-based design would be the foundation of tailor-made molecules and materials toolbox. With such an integrated toolbox at hand the virtual testing concept and application is realized. This challenging task can only be accomplished through a variety of scale bridging methods which


360

A. Mosavi and T. Rabczuk

requires machine learning and optimization combined [4]. Furthermore a great deal of understanding on big data and prediction technologies for microstructure behavior of existing materials, as well as the ability to test the behavior of new materials at the atomic, microscopic and mesoscale is desired to confidently modifying the materials properties [7]. Numerical analysis further allows efficient experiments with entirely new materials and molecules [20]. Basic machine learning technologies such as artificial neural networks [21], and genetic algorithms [9], Bayesian probabilities and machine learning [8], data mining of spectral decompositions [7], refinement and optimization by cluster expansion [20], structure map analysis and neural networks [1], and support vector machines [19], have been recently used for this purpose. Computational materials design innovation to perfect needs to dramatically improve and put crucial components in place. To be precise, data mining, efficient codes, Big data technologies, advanced machine learning techniques, intelligent and interactive MOO, open and distributed networks of repositories, fast and effective descriptors, and strategies to transfer knowledge to practical implementations are the research gaps to be addressed [6]. In fact the current solvers rely only on a single algorithm and address limited scales of the design problems [17]. In addition there is a lack of reliable visualization tools to better involve engineers into the design loop [11]. The absence of robust design, lack of the post-processing tools for multicriteria decision-making, lack of Big data tools for an effective consideration of huge materials database are further research gaps reported in literature [8]. To conclude, the process of computational material design innovation requires a set of up-to-date solvers to cover a wide range of problems. Further problem with the current open-source software toolboxes, reported in [6], is that they require a concrete specification on the mathematical model, and also the modeling solution is not flexible and adaptive. This has been a reason why the traditional computation tools for materials design have not been realistic and as effective. Consequently the vision of this work is to propose an interactive toolbox, where the solver determines the optimal choices via visualization tools as demonstrated in [5]. Ultimately the purpose is to construct a knowledge-based virtual test laboratory to simultaneously optimize the hybrid materials microstructure systems, e.g. textile composites. Whether building atomistic, continuum mechanics or multiscale models, the toolbox can provide a platform to rearrange the appropriate solver according to the problem at hand. Such platform contributes in advancement of innovative materials database leading to innovative materials design with the optimal functionality.

3 LION as a Solver The complex body of information of computational materials design requires the most recent advancements in machine learning and MOO to scale to the complex and multiobjective nature of the optimal materials design problems [10]. From this perspective the materials design can be seen as a high potential research area and a continuous source of challenging problems for LION. In the LION way [3] every individual design task, according to the problem at hand, can be modeled on the basis of the solvers within the toolbox. To obtain a design model the methodology does not ask to specify a model, but it experiments with the current system. The appropriate


Learning and Intelligent Optimization for Material Design Innovation

361

model is created in the toolbox and further is used to identify a better solution in a learning cycle. The methodology is based on transferring data to knowledge to optimal decisions through LION way i.e. a workflow that is referred to as prescriptive analytics [4]. In addition an efďŹ cient Big data application [18] can be integrated to build models and extract knowledge. Consequently a large database containing the properties of the existing and hypothetical materials is interrogated in the search of materials with the desired properties. Knowledge exploits to automate the discovery of improving solutions i.e. connecting insight to decisions and actions [17]. As the result a massively parallelized multiscale materials modeling tools that expand atomistic-simulation-based predictive capability is established which leads to rational design of a variety of innovative materials and applications. A variety of solvers integrated within the LION include several algorithms for data mining, machine learning, and predictive analytics which are tuned by cross-validation. These solvers provide the ability of learning from data, and are empowered by reactive search optimization (RSO) [4] i.e. the intelligent optimization tool that is integrated into the solver. The LION way fosters research and development for intelligent optimization and Reactive Search. Reactive Search stands for the integration of sub-symbolic machine learning techniques into local search heuristics for solving complex optimization problems via an internal online feedback loop for the self-tuning of critical parameters [3, 12]. In fact RSO is the effective building block for solving complex discrete and continuous optimization problems which can cure local minima traps. Further, cooperating RSO coordinates a collection of interacting solvers which is adapted in an online manner to the characteristics of the problem. LIONsolver [4], LIONoso (a non-proďŹ t version of LIONsolver), and Grapheur [5], are the software implementations of the LION way which can be customized for different usage contexts in materials design. These implementations have been used for solving a number of real-life problems including materials selection [18], engineering design [14, 15], computational mechanics [13], and Robotics [16].

4 Textile Composites Optimal Design To evaluate the effectiveness of the LION way the case study of textile composites design with MOO, presented by Milani (2011), is reconsidered using Grapheur. This case study describes a novel application of LION way dealing with decision conflicts often seen among design criteria in composites materials design [18]. In this case study it is necessary to explore optimal design options by simultaneously analyzing materials properties in a multitude of disciplines, design objectives, and scales. The complexity increases with considering the fact that the design objective functions are not mathematically available and designer must be in the loop of optimization to evaluate the Mesomanufacturing Scales of the draping behavior of textile composites. The case study has a relatively large-scale decision space of electrical, mechanical, weight, cost, and environmental attributes. To solve the problem an interactive MOO model is created with Grapheur. With the aid of the 7D visualization graph the designer in the loop formulates and systematically compares different alternatives against the large sets of design criteria to


362

A. Mosavi and T. Rabczuk

tackle complex decision-making task of exploring trade-offs and also designing break-even points. With the designer in the loop, interactive schemes are developed where Grapheur provides a versatile tool for stochastic local search optimization. Once the optimal candidates over the ďŹ ve design objectives preselected, screening the Mesomanufacturing Scales of draping ďŹ gures takes place to identify the most suitable candidate. In this case study the interactive MOO toolset of Grapheur provides a strong user interface for visualizing the results, facilitating the solution analysis, and post-processing (Fig. 1).

Fig. 1. 7D visualization graph for MOO and post-processing: Interactive MOO toolset of Grapheur on exploring trade-offs and simultaneous screening the Mesomanufacturing Scales: the multi-disciplinary property values of candidate materials are supplied from [12].

5 Conclusions Computational material design innovation as an emerging area of materials science requires an adaptive solver to rule a wide range of materials design problems. The LION way provides a suitable platform for developing a computational toolbox for the virtual optimal design and simulation-based optimization of advanced materials to model, simulate, and predict the fundamental properties and behavior of multiscale materials. The proposed solver is a simple yet powerful concept presenting an integration of advanced machine learning and intelligent optimization techniques. With a strong interdisciplinary background the novel application of LION way connects computer science and engineering, and further strengthens the research direction of digital engineering.


Learning and Intelligent Optimization for Material Design Innovation

363

References 1. Artrith, N.H., Alexander, U.: An implementation of artificial neural-network potentials for atomistic materials simulations. Comput. Mater. Sci. 114, 135–150 (2016) 2. Bayer, F.A.: Robust economic Model Predictive Control using stochastic information. Automatica 74, 151–161 (2016) 3. Battiti, R., Brunato, M.: The LION Way: Machine Learning plus Intelligent Optimization. Lionlab, University of Trento, Italy (2015) 4. Brunato, M., Battiti, R.: Learning and intelligent optimization: one ring to rule them all. Proc. VLDB Endow. 6, 1176–1177 (2013) 5. Brunato, M., Battiti, R.: Grapheur: a software architecture for reactive and interactive optimization. In: Blum, C., Battiti, R. (eds.) LION 2010. LNCS, vol. 6073, pp. 232–246. Springer, Heidelberg (2010). doi:10.1007/978-3-642-13800-3_26 6. Ceder, G.: Opportunities and challenges for first-principles materials design and applications to Li battery materials. Mater. Res. Soc. Bull. 35, 693–701 (2010) 7. Fischer, C.: Predicting crystal structure by merging data mining with quantum mechanics. Nat. Mater. 5, 641–646 (2006) 8. Jain, A.: A high-throughput infrastructure for density functional theory calculations. Comput. Mater. Sci. 50, 2295–2310 (2011) 9. Johannesson, G.H.: Combined electronic structure and evolutionary search approach to materials design. Phys. Rev. Lett. 88, 255–268 (2002) 10. Lencer, D.: A map for phase-change materials. Nat. Mater. 7, 972–977 (2008) 11. Mosavi, A.: Decision-making software architecture; the visualization and data mining assisted approach. Int. J. Inf. Comput. Sci 3, 12–26 (2014) 12. Milani, A.: Multiple criteria decision making with life cycle assessment for material selection of composites. Express Polym. Lett. 5, 1062–1074 (2011) 13. Mosavi, A., Vaezipour, A.: Reactive search optimization; application to multiobjective optimization problems. Appl. Math. 3, 1572–1582 (2012) 14. Mosavi, A.: A multicriteria decision making environment for engineering design and production decision-making. Int. J. Comput. Appl. 69, 26–38 (2013) 15. Mosavi, A.: Decision-making in complicated geometrical problems. Int. Comput. Appl. 87, 22–25 (2014) 16. Mosavi, A., Varkonyi, A.: Learning in Robotics. Int. J. Comput. Appl. 157, 8–11 (2017) 17. Mosavi, A., Rabczuk, T., Varkonyi-Koczy, A.R.: Reviewing the novel machine learning tools for materials design. In: Luca, D., Sirghi, L., Costin, C. (eds.) INTER-ACADEMIA 2017: Recent Advances in Technology Research and Education. Advances in Intelligent Systems and Computing, vol. 660, pp. 50–58. Springer, Cham (2018). doi:10.1007/978-3319-67459-9_7 18. Mosavi, A., et al.: Multiple criteria decision making integrated with mechanical modeling of draping for material selection of textile composites. In Proceedings of 15th European Conference on Composite Materials, Venice, Italy (2012) 19. Saito, T.: Computational Materials Design, vol. 34. Springer Science & Business Media, Heidelberg (2013) 20. Stucke, D.P., Crespi, V.H.: Predictions of new crystalline states for assemblies of nanoparticles. Nano Lett. 3, 1183–1186 (2003) 21. Sumpter, B.G., Noid, D.W.: On the design, analysis, and characterization of materials using computational neural networks. Annu. Rev. Mater. Sci. 26, 223–277 (1996)


Turn static files into dynamic content formats.

Create a flipbook
Issuu converts static files into: digital portfolios, online yearbooks, online catalogs, digital photo albums and more. Sign up and create your flipbook.