In This Issue
Spring Bridge on the US Metals Industry: Looking Forward
March 29, 2024 Volume 54 Issue 1
In this issue of The Bridge, guest editors Greg Olson and Aziz Asphahani have assembled feature articles that demonstrate how computational materials science and engineering is leading the way in the deployment of metallic materials that meet increasingly advanced design specifications.

Harnessing the Power of AI in Materials Digital Transformation: A Synergistic Hybrid Approach

Monday, April 8, 2024

Author: Jiadong Gong

The synergy between AI and materials science holds immense promise for the future of materials engineering.

The quest for novel materials, alongside their optimization and adoption, has perennially fueled the engine of innovation. Traditionally, the development of materials relied heavily on a trial-and-error approach, characterized by extensive experimentation. However, this method often proved to be lengthy, arduous, and costly. That narrative began to shift with the emergence of computational methods and tools like computational thermodynamics (CALPHAD) (Saunders 1998) and density functional theory (DFT) (Kohn 1999), ushering in the era of computational materials design. Greg Olson pioneered this approach in a seminal paper (Olson 1997), setting the stage for a new paradigm in materials design and development. The value of the CALPHAD-grounded approach, trademarked as Materials by Design®, is that it allows us to use our existing predictive mechanistic knowledge in a quantitative and system-specific way (Olson et al. 2014). The efficacy of this approach was subsequently underscored by the accomplishments of companies such as QuesTek Innovations, SpaceX, Tesla, and Apple across diverse sectors including energy, aerospace, automotive, and electronics (Kuehmann 2024; Warren 2024).

The early triumphs of computational methodologies in expediting material innovation captured the attention of the US government, culminating in the inception of the Materials Genome Initiative in 2011. This ambitious venture aimed to halve the time and cost of the conventional ten-to-twenty-year cycle of material creation and deployment by funding and developing foundational databases and computational tools. As the academic and industrial realms increasingly embraced this initiative, a plethora of tools emerged, including various science-based modeling tools across various time and length scales. This proliferation led to the widespread adoption of integrated computational materials engineering (ICME), a concept that was broached and demonstrated in an earlier NRC study and the DARPA-AIM program (NRC NMAB 2004, NRC NMAB 2008). In the hands of leading corporations, this technology has now reduced the full materials design and deployment cycle to under two years.

In tandem with the expansion of computational tools and databases came a surge in computational power that is ready for use. This prompted the proposition of a “­digital twin” for materials, envisioned to obviate the necessity for physical trial-and-error experiments. The concept posited that all such experiments could be conducted in a digital realm, aligning with the broader ­narrative of digital transformation in materials engineering. The ascendancy of artificial intelligence and machine learning (AI/ML), epitomized by advancements in language processing, computer vision, optimization, and forecasting, has propelled this idea further (Choudhary et al. 2022; Gomes et al. 2019). The recent emergence of GPT (Vaswani et al. 2017), a large language model (LLM) from OpenAI, captured significant attention from both the research and business communities, marking a ­pivotal moment in the advent of generative AI (GenAI). This new wave of AI holds promises in promoting the generation of new materials, although the reports surrounding fully AI-created materials remain nascent and challenges abound.

Gong Fig 1.gifExamining the Challenges

The anticipation of AI as a catalyst for materials innovation is indeed burgeoning. A few have spotlighted promising outcomes in the domain of robotic and autonomous experimentation (Pyzer-Knapp et al. 2022). Particularly, self-driving laboratories have made strides in new thin-film discoveries (MacLeod et al. 2020), hinting at the burgeoning potential of AI in revolutionizing the ­materials engineering landscape. Notably, researchers have been delving into “materials discovery” (Gomes et al. 2019; Merchant et al. 2023), allowing algorithms to traverse the expansive space of material compositions, with hopes of “discovering” new materials with coveted properties. However, the fruits of these endeavors haven’t quite lived up to the initial expectations. Several factors underpin this shortfall, and below are some notable challenges:

1. Data Deficiency: Quantity and Quality Matters

The march towards an AI-first or data-driven approach in materials engineering has been championed by many, thanks to the growth in computing power and the emergence of sophisticated AI/ML algorithms (Chollet 2019). Unlike domains such as social economic studies or ­digital media, where data is abundant, materials engineering often grapples with data scarcity and sparsity (Wang et al. 2020). The available data needed for direct training of comprehensive material models is just not adequate. Complicating matters further, a significant chunk of valuable data resides as proprietary within the private sector. Moreover, the available data often comes with various errors and outliers. In addition, a well-trained AI/ML model promises efficiency after deployment, but the ­preparatory steps, from data acquisition, cleaning, and curation to training and validation, often make this approach much slower and inefficient than science-based modeling.

This new wave of AI holds promises in promoting the generation of new materials, although the reports surrounding fully AI-created materials remain nascent and challenges abound.

2. Algorithmic Problem: Navigating the Black Box

The quality of data and inherent biases within AI/ML algorithms markedly influence their performance and reliability. The usual black-box nature of most AI algorithms poses more challenges in understanding their modeling processes. This understanding is ­critical for validation and trust, especially in the context of ­materials engineering, where erroneous predictions can have substantial consequences. A striking illustration is the comparison between the interpolation-centric nature of AI models versus the extrapolation capabilities of physics-based models. AI models may falter, generating ­unrealistic predictions like negative strength values for materials, underscoring the necessity for particular accuracy and reliability in modeling outcomes.

3. Complexity of Materials: Hierarchical Systems

Applied engineering materials are usually complicated systems with various subsystems at different length scales; they aren’t merely waiting to be discovered. Apart from some success in unearthing new chemical compounds, the creation of new materials necessitates deliberate designs. Practical material applications exhibit complex interlinks among process, structure, properties, and performance (PSPP), resulting in a hierarchical structure of multiscale systems, with a mixed state of stable or metastable, equilibrium or non-equilibrium. This complexity underscores the indispensable role of domain expertise in developing and deploying meaningful AI solutions in this field.

Despite the challenges of direct applications, there are potential opportunities for the role of AI in materials engineering, especially within the context of digital transformation.

Certain computational methods, like DFT, are suitable for ML techniques for synthetic data generation and exploring specific chemical compound spaces (CCS). DFT-based ML models have shown promise in efficiency and scalability (Huang et al. 2023), hinting at the potential for experimental planning within self-driving laboratories. Yet, the success in identifying suitable compounds for certain property combinations hasn’t translated to broader material development applications. The crux lies in the multimode, multi-structured nature of almost all practical engineering materials, which is one key difference between materials development and chemical engineering.

Where are the Opportunities?

While raw data-driven AI techniques have had success in identifying new chemical compounds, far more success has been found in the direct application of mechanistic knowledge in the efficient parametric design of complex multiphase materials, aided by “genomic-level” fundamental CALPHAD databases. Notably, this has taken “clean sheet” designs of aircraft landing gear steels all the way to flight qualification (NIST 2016; NRC 2012). We next explore the opportunity of hybrid approaches within the framework of “structured hybrid modelling” (Bhutani 2006), which efficiently integrates mechanistic and empirical modeling.

Despite the challenges of direct applications, there are potential opportunities for the role of AI in ­materials engineering, especially within the context of digital transformation. As previously noted, the lack of domain­specific data and materials science expertise has resulted in significant gaps in the AI-centric approaches discussed in the preceding section. Moreover, this deficit in domain knowledge underscores why AI approaches, when purely data-driven and devoid of underlying scientific comprehension, demand a high volume of quality data and further exacerbate the data need. Overcoming this hurdle could markedly accelerate our journey towards a new era of engineering materials design and deployment.

Leveraging well-established science-based models can drastically trim the data requirement. A handful of critical calibration data points can yield highly accurate outcomes. These models are predominantly mechanistic and in analytic forms, with superior speed and efficiency in predicting PSPP relationships, without the cumbersome process of data collection and training.

Yet, some science-based models are stochastic or necessitate extensive simulation. While these models excel in prediction with scientific soundness, integrating them with AI/ML into hybrid models could potentially bring improved efficiency. The principles of ICME require the interconnection of models across various length and time scales, compelling materials modelers or designers to judiciously select model combinations that optimize overall efficiency. For example, atomistic simulations for materials are very powerful and accurate with high-­fidelity atomic interactions; however, they are greatly limited by the large computational cost. Recent advancements in machine-learning interatomic potentials (MLIAP) have given access to atomistic simulations that reach similar accuracy levels but are orders of magnitude faster, thus dramatically widening the spectrum of materials systems that can be simulated with high physical fidelity ­(Deringer et al. 2019). Additionally, in scenarios where linkages within a particular PSPP system remain elusive and well-established scientific models are absent, AI/ML models could serve as valuable additions to bridge these gaps.

We have already recognized the success of AI in discovering new compounds within expansive CCS, as high­lighted by the recent work by Google DeepMind (­Merchant et al. 2023). However, the intricacy of engineering materials, underscored by their multiscale structures, renders the discovery of new materials via AI a challenging task. Nonetheless, the compounds discovered through AI could be promising candidates for constituents at different structural levels, enabling ambitious designs. In a multiscale materials system, a series of design choices exist for employing particular phases for diverse purposes, such as strengthening precipitates, grain-­pinning particles, and inoculants, among others. Employing AI to fill the gaps in missing models of the PSPP linkage and to identify coveted microstructure constituents in the ICME design chart could significantly enhance overall materials design. Furthermore, the deployment of well-established AI algorithms for smart search and optimization could further fine-tune this ­process’s efficiency. Below are two case studies to demonstrate some of these aspects.

Case Study 1: High-Temperature Nb Alloy Design

The aspiration to deploy Nb-based alloys as a viable upgrade for Ni-based superalloys is rooted in their potential for superior performance in high-temperature applications, such as rocket nozzles and next-generation turbines. However, realizing this goal requires over­coming ­formidable design hurdles, including achieving high ­specific strength, creep resistance, and oxidation resistance at elevated temperatures while preserving ­ductility at lower temperatures. Additionally, the requirement for alloy coatings to ensure compatibility with coating ­materials further complicates the design space.

QuesTek Innovations has employed a blend of computational methodologies to address some of these chal­lenges. One of the key tasks was to delineate a strengthening strategy for the new alloy. Beyond the conventional solid solution strengthening from alloy elements in the Nb matrix, two other potent mechanisms were explored: precipitation strengthening and grain refinement through dispersion. For effective high-­temperature precipitation strengthening, the chosen precipitate should be ­coherent with the Nb matrix (i.e., low crystal mismatch) and demonstrate robust coarsening resistance at operating temperatures (i.e., high-temperature thermal stability). Typical coherent, ordered bcc-type precipitate systems include the B2 (stoichiometry AB) and L21 ­(stoichiometry A2BC) structures. Similarly, the design of a fine grain size is achieved through the formation of a small fraction of cubic MX (X=O, C, N) oxycarbide grain pinning dispersions. The strategy is to leverage the typical level of interstitial elements in Nb alloy to design a stable MX dispersion with suitable coarsening resistance and solvus temperature, ensuring refined grain size across various processing stages. However, employing commonly used phases to meet these stringent, sometimes conflicting constraints posed a significant challenge.

The integration of AI in engineering materials design heralds a new frontier of innovation, yet it could also bring along a spectrum of risks that may necessitate regulatory measures
to ensure safe and
responsible deployment.

A potential solution is the multicomponent ­concept—multiple substitutional elements within the same ­sublattice—to exploit the high entropy effect for phase stabilization. While seasoned materials designers could search available CALPHAD databases for well-assessed compound spaces, data pertaining to these multi­component chemical spaces largely remained elusive. Employing hybrid machine learning and DFT ­approaches (Choudhary 2022) for a high-throughput search within a vast CCS, potent phases were identified that satisfied all constraints, specifically a multicomponent L21 strengthening precipitate and the MX oxycarbide grain pinner.

Low-temperature ductility, another crucial factor, ­hinges significantly on the precise design of the ductile-brittle transition temperature (DBTT). A traditional parametric DBTT model grounded in the “master curve” framework of Odette, employing a universal hardness dependence (Odette and Lucas 2001), could delineate contributions from different alloying elements for the solid solution of the niobium matrix. Yet, the ­endeavor to develop an accurate DBTT model within a high-­dimensional multicomponent space was restricted by a dearth of relevant data for effective regression. QuesTek gathered a specific dataset encapsulating DBTT data for twenty-six multicomponent Nb alloys from available literature and internal datasets. Utilizing a statistical learning ­method, QuesTek was able to identify effective ­materials descriptors from possibly correlated feature spaces, enabling accurate DBTT prediction for multicomponent Nb alloys with very limited data.

With the proven Materials by Design® methodology under the ICME framework, QuesTek successfully engineered a novel Nb alloy that met the stringent design requirements. Some of the design highlights are illustrated in figure 1.

Case Study 2: Shape-Memory Materials

Shape-memory materials, with their high functional potential, have been employed in a myriad of applications. The shape-memory effect primarily hinges on the martensitic transformations, which are diffusionless, displacive phase transitions and can yield significant engineering benefits such as hardening, toughening, and shape-memory properties across various material systems, from alloys to ceramics. The martensitic transformation induces large strains, creating a complex issue as the transforming region coexisting with the surrounding untransformed matrix, thus leading to a considerable mismatch. In ceramic systems, this behavior is often associated with deleterious bond-breaking events reaching the material’s elastic limit and the generation of transformation hysteresis. In what follows, a recent advancement in ceramics is examined to shed light on the transferable design methodology utilizing a combination of different modeling techniques.

Gong Fig 2.gif

Pang and colleagues (2022) utilized a multifaceted modeling approach, interlinking machine learning, computational thermodynamics, and lattice engineering to predict shape-memory characteristics of new ZrO2-based compositions (see figure 2). Zirconia ceramics exhibit a ­martensitic phase transformation permitting large strains, thus emerging as candidates for shape-memory and superelastic applications at high temperatures. Like other martensitic materials, the transformation strain can be engineered by alloying to yield a more commensurate transformation with reduced hysteresis. Yet, “­lattice engineering” in ­zirconia is further complicated by additional physical constraints such as managing a large transformation volume change and achieving transformation temperatures high enough to evade kinetic barriers. Due to data scarcity for materials of interest, elements of data science, including supervised machine learning, were introduced to navigate a complex multidimensional search space. The multi-objective optimization outcomes directed targeted experiments, which subsequently led to a cracking-resistant polycrystalline martensitic zirconia ceramic with record low thermal hysteresis.

The outcome was a new zirconia composition with a low hysteresis of 15 K, showcasing about five times less than the best values reported thus far. This revelation implies that zirconia ceramics can exhibit hysteresis ­values comparable to those of widely deployed shape-memory alloys, thereby opening avenues for their application as high-temperature shape-memory materials. Compared to other reports of a pure data-driven AI approach for similar shape-memory designs (Trehern et al. 2022), the methodologies deployed herein that utilize different modes of models in both ­parallel and serial fashion can be readily transferred to the design of shape-memory alloys or other material systems upon careful modification and calibration.

Gong Fig 3.gif

Smart Decision-Making Drives Innovations

A simplistic yet factual interpretation of decision-making activities can be seen in figure 3. Here, data in its myriad forms—structured or unstructured, measured or synthesized, human-observed or sensor-derived—serves as the foundation. All terrestrial entities, whether humans or machines, process received data with prior knowledge through intrinsic models to generate decisions or actions. The essence of these models, their ability to generate the right decisions from the data, embodies the coveted trait of intelligence, be it artificial or natural.

The evolution of these models can be aptly segmented into four progressive categories: descriptive, predictive, prescriptive, and generative. The advent of large language models (LLMs) like ChatGPT has ignited substantial buzz around generative artificial intelligence (GenAI). Despite the appeal of a super GenAI tool capable of generating all kinds of recommendations and decisions, the importance of ensuring that models in critical engineering applications accurately predict outcomes and recommend appropriate solutions remains paramount.

Presently, science-based models and simulations remain the workhorses in accelerating engineering innovations. Yet, as elucidated in previous sections, AI techniques can significantly enhance this decision-­making process from all aspects across data, models, and ­materials domain knowledge. AI tools could essentially enable engineers from diverse fields to articulate intricate simulation needs in natural language and construct complex domain-­specific materials models. Such simplification could potentially lower the domain expertise threshold significantly. A notable example is MIT researchers who, from mechanical engineering practice, combined ICME and machine learning techniques in a hybrid approach for the inverse design of additively manufacturable high-strength aluminum alloys, adeptly navigating a complex high-dimensional materials space while adhering to all constraints (Taheri-Mousavi et al. 2024). As emerging AI techniques mature and proliferate, anticipation surges around their potential to augment, automate, and optimize more facets of the decision-making process. Interweaving these advancements with science-based modeling and simulation could markedly elevate ­materials selection and design efficiency, reduce development and ­deployment time and cost, and nurture cross-disciplinary collaborations.

Risks and Regulatory Aspects

The integration of AI in engineering materials design ­heralds a new frontier of innovation, yet it could also bring along a spectrum of risks that may necessitate regulatory measures to ensure safe and responsible deployment. The dependency on data, in particular, raises concerns regarding data privacy and security, especially when sensitive or proprietary information is involved. Ensuring the confidentiality and integrity of data is paramount to mitigate the risks associated with data breaches or misuse. The potential impact of biases and inaccuracies generated by lower-quality data can be disastrous for materials applications, and the black-box nature of AI algorithms poses additional challenges in validating and trusting their predictions. Determining accountability in cases of adverse outcomes resulting from AI-driven decisions in materials engineering could be a complex legal and ethical challenge. Clear liability frameworks are crucial to addressing potential disputes and to ensuring adherence to ethical standards.

The creation of innovative materials through AI is indeed the crown jewel of hopes, but it also poses questions regarding who owns intellectual property rights. Establishing clear guidelines on this issue of ownership is necessary to foster innovation while ensuring legal clarity. Developing standards and best practices can help harmonize the regulatory landscape and ensure the safe adoption of AI. Given the global nature of both AI and materials engineering, international collaboration is vital to address the associated risks and foster a conducive environment for leveraging the benefits of AI in ­materials engineering to address global issues, such as climate change and sustainability.

A Synergistic Future

The narrative of a “digital twin” for materials, once a far-fetched notion, is gradually becoming a tangible reality, fueled by the relentless advancements in computational power and AI technologies. The integration of AI/ML with established ICME technologies presents a new frontier in the digital transformation of materials innovations, ushering in a synergistic hybrid approach. This combination expedites the quest for novel materials and optimization strategies, blending the rigor of scientific models with the adaptive capability of AI techniques. This synergy is poised to significantly reduce the time and resources conventionally required in the materials engineering domain, fostering a quicker transition from conceptualization to deployment. The examples illustrated in this article showcase the early tangible successes where a congruous integration can amplify the overall efficacy, and they also underscore the immense potential awaiting realization.

As we stand on the cusp of this juncture, the future beckons with the promise of accelerated advancements, rendering previously insurmountable challenges surmountable. The maturation of AI techniques, coupled with a growing understanding and adoption of a hybrid approach, could potentially redefine the landscape of materials engineering. The journey ahead, though laden with challenges, emanates a beacon of promise for a future where the synergy between AI and materials science catalyzes a new era of materials innovation, driving global progress in myriad applications.

Acknowledgement

The author acknowledges the support from ARPA-E ULTIMATE Program, DARPA SIMPLEX Program, and the collaboration from NIST CHiMaD center.

References

Bhutani N, Rangaiah GP, Ray AK. 2006. First-principles, data-based, and hybrid modeling and optimization of an industrial hydrocracking unit. Industrial & Engineering Chemistry Research 45(23):7807–7816.

Chollet F. 2019. On the measure of intelligence. arXiv preprint arXiv:1911.01547.

Choudhary K, DeCost B, Chen C, Jain A, Tavazza F, Cohn R, Park CW, Choudhary A, Agrawal A, Billinge SJL, and 3 others. 2022. Recent advances and applications of deep learning methods in materials science. npj Computational Materials 8(1):59.

Deringer VL, Caro MA, Csányi G. 2019. Machine learning interatomic potentials as emerging tools for materials science. Advanced Materials 31(46):1902765.

Gomes CP, Selman B, Gregoire JM. 2019. Artificial intelligence for materials discovery. MRS Bulletin 44:538–544.

Huang B, Von Rudorff GF, Von Lilienfeld OA. 2023. The ­central role of density functional theory in the AI age. ­Science 381(6654):170–175.

Kohn W. 1999. Nobel Lecture: Electronic structure of matter—wave functions and density functionals. Reviews of Modern Physics 71:1253.

Kuehmann CJ. 2024. Concurrent Design of materials and ­systems. The Bridge 54(1):38–44.

MacLeod BP, Parlane FGL, Morrissey TD, Hase F, Roch ML, Dettelbach KE, Moreira R, Yunker LPE, Rooney MB, Deeth JR, and 10 others. 2020. Self-driving laboratory for accelerated discovery of thin-film materials. Sciences Advances 6(20):eaaz8867.

Merchant A, Batzner S, Schoenholz SS, Aykol M, Cheon G, Cubuk ED. 2023. Scaling deep learning for materials dis­covery. Nature 624:80–85.

National Institute of Standards and Technology (NIST). 2016. Materials Innovation Case Study: QuesTek’s Ferrium® M54® Steel for Hook Shank Application. Nexight Group and Energetics, Inc.

National Research Council (NRC). National Materials Advisory Board. 2004. Accelerating Technology Transition: Bridging the Valley of Death for Materials and Processes in Defense Systems. Washington, DC: The National Academies Press.

NRC. National Materials Advisory Board. 2008. Integrated ­Computational Materials Engineering: A Transformational Discipline for Improved Competitiveness and National ­Security. Washington, DC: The National Academies Press.

NRC. 2012. Application of Lightweighting Technology to ­Military Aircraft, Vessels, and Vehicles. Washington, DC: The National Academies Press.

National Science And Technology Council (NSTC). 2011. Materials Genome Initiative for Global Competitiveness. Washington, DC.

Odette GR, Lucas GE. 2001. Embrittlement of nuclear reactor pressure vessels. JOM 53:18–22.

Olson GB. 1997. Computational design of hierarchically structured materials. Science 277:1237–42.

Olson GB, Kuehmann CJ. 2014. Materials genomics: From CALPHAD to flight. Scripta Mater 70:25–30.

Pang EL, Olson GB, Schuh CA. 2022. Low-hysteresis shape-memory ceramics designed by multimode modelling. Nature 610:491–495.

Pyzer-Knapp EO, Pitera JW, Staar PWJ, Takeda S, Laino T, Sanders DP, Sexton J, Smith JR. 2022. Accelerating ­materials discovery using artificial intelligence, high performance computing and robotics. npj Computational Materials 8:84.

Saunders N, Miodownik AP. 1998. CALPHAD—Calculation of Phase Diagrams, A Comprehensive Guide. Oxford, England: Elsevier.

Taheri-Mousavi SM, Xu M, Olson GB, Hart AJ, and 6 others. 2024. Additively manufacturable high-strength aluminum alloys exploiting nano-metastability via rapid solidification. Nature, manuscript under review.

Trehern W, Ortiz-Ayala R, Atli KC, Arróyave R, Karaman I. 2022. Data-driven shape memory alloy discovery using artificial intelligence materials selection (AIMS) framework. Acta Materialia 228:117751.

Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, Kaiser Ł, Polosukhin I. 2017. Attention is all you need. In: Advances in Neural Information Processing Systems 30. Guyon I, Von Luxburg U, Bengio S, Wallach H, Fergus R, Vishwanathan S, Garnett R, eds. San Diego, California: Neural Information Processing Systems Foundation, Inc. (NeurIPS).

Wang AYT, Murdock RJ, Kauwe SK, Oliynyk AO, Gurlo A, Brgoch J, Persson KA, Sparks TD. 2020. Machine learning for materials scientists: An introductory guide toward best practices. Chemistry of Materials 32(12):4954–4965.

Warren J. 2024. The materials genome initiative and the metals industry. The Bridge 54(1):25–29.

 

About the Author:Jiadong Gong is chief technology offcer and chief information officer, QuesTek Innovations.