Can Artificial Intelligence Accelerate Technological Progress? Researchers' Perspectives on AI in Manufacturing and Materials Science

The gist Artificial intelligence (AI) was primarily used for modeling of materials and manufacturing processes, facilitating cheaper and more rapid search of design spaces for materials and manufacturing processes alike, with benefits including cost, time, and computation savings in technology development.

How it works

The research explored how AI can accelerate technological progress by investigating whether it improves the breadth, reliability, or cost at which new technological alternatives can be tested and whether it is taking over research tasks previously performed by humans. The core question of whether and how AI can facilitate technological progress has not been answered, requiring an investigation into how AI applications and affordances affect fundamental components of the process of technological change. The study framed the question within longstanding evolutionary accounts of technological change, asking whether AI improves the breadth, reliability, or cost at which new technological alternatives can be tested and whether it is taking over research tasks previously performed by humans.

Key Findings on AI Applications

The results from 32 interviews with U.S.-based academic manufacturing and materials sciences researchers found that AI was primarily used for modeling of materials and manufacturing processes, facilitating cheaper and more rapid search of design spaces for materials and manufacturing processes alike. Benefits included cost, time, and computation savings in technology development. AI/ML tools were found to be useful in two broad categories of novel modeling opportunities: first, surrogate models for preexisting physics-based models, which can yield similar results at lower computational cost, and second, models trained on input and output data for phenomena lacking robust theoretical models, referred to as phenomenological modeling.

Benefits of AI/ML in Research

Interviewees discussed several benefits of AI/ML use in research, including the Reduction in accuracy/cost/speed trade-off in research, especially computer modeling, where AI can produce similar results to conventional, physics-based simulations at lower computational cost. Other benefits included Exploring larger design spaces and the ability to Address of previously unsolvable problems, such as identifying human-unidentifiable patterns or phenomena. AI/ML can also be used to Planning experimental campaigns and can reduce the computation time required to achieve a given result, saving on electricity, hardware use, and, often, time.

Downsides and Limitations

Despite the benefits, interviewees noted significant downsides. The most commonly discussed issue was Accuracy weaknesses, where AI/ML tools can Predict poorly outside regions of dense, high-quality training data, as they are limited to interpolation rather than extrapolation. Furthermore, there were concerns about Interpretability weaknesses, as the bounds of accuracy can be unclear and it is difficult to judge their situational utility or repair errors. A core scientific concern was that AI/ML lacks the ability to develop novel scientific theory, as it is primarily focused on interpolation within known, trained data sets.

Conclusion and Implications

The study suggests that AI/ML is more another tool in the research and development toolkit than a radical transformation of the innovation process. Conventional experimental and theory development methods remain essential, as AI/ML is best suited for sustaining innovation within existing technological paradigms. The findings imply that investment in AI/ML should be integrated within a robust portfolio of support and funding for more conventional methods and skills development. Ultimately, reliable, general, and context-sensitive inference about AI’s uses and effects in science requires a situation of empirical findings within durable theoretical understanding of the intertwined processes of scientific and technological change.

The results suggest that research strategists and policymakers should be cautious about claims of AI/ML fundamentally revolutionizing science and technological development. AI/ML is best thought of as adding another “rung” to the ladder of vicarious trial in some areas, providing a method to predict the viability of different technological alternatives at lower cost. This utility is important and valuable, but simultaneously contingent and limited. AI/ML can certainly benefit research and development, but they require skilled users, and integration within robust systems of more conventional theory development and technological trial. Research funders and policymakers should regard investment in AI/ML infrastructure as a complement to conventional elements of the research enterprise rather than an independent substitute. The findings suggest that AI/ML may be a sustaining, thus entrenching, innovation to the process of sustaining innovation itself, circumventing swaths of laborious empirical exploration in which investigators might stumble over unforeseen and disruptive opportunities. This is not to say that AI/ML has unlimited utility; many interviewees indicated that AI/ML-based models are most useful for interpolation of phenomena within well-mapped physical domains. While AI/ML tools bear potential to decrease the resource intensity of at least some research, several interviewees noted that the reverse can occur, as training and data acquisition are often expensive and labor-intensive processes. Research managers and sponsors should ensure attention to whole workflow operation when funding and supporting the adoption of AI/ML technologies. The study suggests that MMS researchers will require a diverse and flourishing research ecosystem including all these subcomponents to maintain technological progress, indeed to maintain the utility of AI/ML itself. This is not, however, to say that AI/ML has unlimited utility; many interviewees indicated that AI/ML-based models are most useful for interpolation of phenomena within well-mapped physical domains and have not (as yet) demonstrated the capability to develop novel, powerful theory capable of major predictive extrapolations. The prospect does not seem likely at present. This is not to say that AI/ML has unlimited utility; many interviewees indicated that AI/ML-based models are most useful for interpolation of phenomena within well-mapped physical domains and have not (as yet) demonstrated the capability to develop novel, powerful theory capable of major predictive extrapolations. The prospect does not seem likely at present. This is not to say that AI/ML has unlimited utility; many interviewees indicated that AI/ML-based models are most useful for interpolation of phenomena within well-mapped physical domains and have not (as yet) demonstrated the capability to develop novel, powerful theory capable of major predictive extrapolations. The prospect does not seem likely at present. This is not to say that AI/ML has unlimited utility; many interviewees indicated that AI/ML-based models are most useful for interpolation of phenomena within well-mapped physical domains and have not (as yet) demonstrated the capability to develop novel, powerful theory capable of major predictive extrapolations. The prospect does not seem likely at present. This is not to say that AI/ML has unlimited utility; many interviewees indicated that AI/ML-based models are most useful for interpolation of phenomena within well-mapped physical domains and have not (as yet) demonstrated the capability to develop novel, powerful theory capable of major predictive extrapolations. The prospect does not seem likely at present. This is not to say that AI/ML has unlimited utility; many interviewees indicated that AI/ML-based models are most useful for interpolation of phenomena within well-mapped physical domains and have not (as yet) demonstrated the capability to develop novel, powerful theory capable of major predictive extrapolations. The prospect does not seem likely at present. This is not to say that AI/ML has unlimited utility; many interviewees indicated that AI/ML-based models are most useful for interpolation of phenomena within well-mapped physical domains and have not (as yet) demonstrated the capability to develop novel, powerful theory capable of major predictive extrapolations. The prospect does not seem likely at present. This is not to say that AI/ML has unlimited utility; many interviewees indicated that AI/ML-based models are most useful for interpolation of phenomena within well-mapped physical domains and have not (as yet) demonstrated the capability to develop novel, powerful theory capable of major predictive extrapolations. The prospect does not seem likely at present. This is not to say that AI/ML has unlimited utility; many interviewees indicated that AI/ML-based models are most useful for interpolation of phenomena within well-mapped physical domains and have not (as yet) demonstrated the capability to develop novel, powerful theory capable of major predictive extrapolations. The prospect does not seem likely at present. This is not to say that AI/ML has unlimited utility; many interviewees indicated that AI/ML-based models are most useful for interpolation of phenomena within well-mapped physical domains and have not (as yet) demonstrated the capability to develop novel, powerful theory capable of major predictive extrapolations. The prospect does not seem likely at present. This is not to say that AI/ML has unlimited utility; many interviewees indicated that AI/ML-based models are most useful for interpolation of phenomena within well-mapped physical domains and have not (as yet) demonstrated the capability to develop novel, powerful theory capable of major predictive extrapolations. The prospect does not seem likely at present.

1. Surrogate modeling for physics-based models can be used to yield similar results at lower computational cost by training an AI/ML model on inputs and outputs from a computationally intensive physics-based model, allowing for speed improvement of up to ten to the fifth (AI02).

2. AI/ML tools can be used as surrogate models for preexisting physics-based models or to address entire phenomena where no robust theoretical models exist, effectively allowing the researcher to surrender the actual task of theory formulation to the model.

3. AI/ML can facilitate planning experimental campaigns by acting as a guide through a design space, such as machine learning-guided walks through a design of experiments landscape (AI09), helping researchers select promising options for real-world trial.

4. AI/ML can automate data preprocessing and preliminary interpretation, which allows for the reduction in accuracy/cost/speed trade-off in research by reducing computation time and saving research labor (Table 4).

5. AI systems can be trained to map sensor readings to process health or machine status, enabling researchers to identify human-unidentifiable patterns or phenomena and model unknown relationships without explicit, human-designed theory.

Artificial intelligence (AI) raises expectations of substantial increases in rates of technological progress, but such anticipations are often not connected to detailed ground-level studies of AI use in innovation processes. Accordingly, it remains unclear how and to what extent AI can accelerate innovation. To help to fill this gap, we explore and assess results from 32 interviews with U.S.-based academic manufacturing and materials sciences researchers experienced with AI and machine learning (ML) techniques. We found that AI was primarily used for modeling of materials and manufacturing processes, facilitating cheaper and more rapid search of design spaces for materials and manufacturing processes alike. Benefits included cost, time, and computation savings in technology development. However, AI/ML tools were unreliable outside design spaces for which dense data were already available; they required skilled and judicious application in tandem with older research techniques; and concerns were raised about the potential to detrimentally circumvent opportunities for disruptive theoretical advancement. Based on these results, we suggest there is reason for optimism about acceleration in sustaining innovations through the use of AI/ML; but that support for conventional empirical, computational, and theoretical research is required to maintain the likelihood of further disruptive advances in manufacturing and materials.