Towards Creative Generative Models for Scientific Discovery

Scientific discovery and automated design are promising application areas for generative foundation models. Complex cyberphysical systems (CPS) such as drones and underwater vehicles present a socio-economically impactful and intellectually challenging domain for automating discovery and design. These systems are often multiphysics and their complexity far exceeds the current applications areas such as material design, proposing new culinary recipe, and protein synthesis. Traditional workflows for cyberphysical systems create detailed digital models which can be evaluated by physics simulators to narrow the search space before creating physical prototypes. A major bottleneck of this approach is that the simulators are often computationally expensive and slow. Recent advancements in machine learning offer the possibility to accelerate these pipelines using surrogate models. Another big challenge is the dearth of data which renders data-hungry deep learning methods not directly applicable. In our recent work, we have developed a neuro-symbolic approach that bootstraps learning from zero or very few examples and creates diverse set of high-performing designs. We created and publicly released AircraftVerse (https://github.com/SRI-CSL/AircraftVerse), an aerial vehicle design dataset. 

Figure 1: AircraftVerse contains 27,714 diverse air vehicle designs - the largest corpus of engineering designs with this level of complexity. The corpus has designs with very different structures and performance. (Left) Each design in AircraftVerse in the dataset comprises a symbolic design tree describing topology, propulsion. (Right) We train an LLM using a curriculum of training tasks ultimately creating subsystem, battery subsystem, and other design details, a model that can generate designs conditions on properties of flight dynamics a STEP model data, an STL file and a 3D point cloud (hover time, maximum speed, etc.).

While these designs are diverse, are these also creative? Computational creativity has been studied in niche domains such as game design. But a formal definition of creativity (and the related concept of novelty) is difficult to construct broadly or even just for cyberphysical systems. In this talk, we will describe our initial efforts for the definition and assessment of creativity. Outside of computational and information science, creativity has been studied from the perspective of cognitive science, psychology, and sociology. Such definitions often characterize the properties of a creative artifact, such as being unusual, being useful, rejecting previously held ideas, and providing clarity. But these do not provide an operational definition that can be used to evaluate creativity or bias generative models to produce creative designs. We build our framework on the multistage characterization of human creativity where creativity is modularized into several (not necessarily sequential) stages of problem definition, knowledge acquisition, information gathering, incubation or synthesis of acquired knowledge, idea generation, idea combination, selection, and representation. We will discuss how these stages, corresponding to divergent or convergent phases of creativity, can be used to evaluate and improve creativity of generative models.

Dr. Susmit Jha is a Technical Director in Computer Science Laboratory at SRI, where he leads the research group on Neuro-Symbolic Computing and Intelligence. Dr. Jha completed his Ph.D. in Computer Science from UC Berkeley in 2011, where his thesis work on “Automated Synthesis Using Structurally Constrained Induction and Deduction” was supported by Berkeley Fellowship and was awarded the Leon O Chua Award. His program synthesis work influenced the development of the FlashFill feature in Excel. Before joining SRI, Dr. Jha was at Intel Labs and Raytheon Technologies Research Center at Berkeley. At Intel, Dr. Jha’s research received a Division Recognition Award in 2012 and Research Technology Scoping Award in 2014. He received the 10-year Most Influential Paper award at IEEE/ACM ICSE 2020. He has published over 90 peer-reviewed publications with over 3700 citations in AI, ML, Formal Methods, and Automated Reasoning venues such as NeurIPS, ICLR, ICML, CVPR, AAAI, IJCAI, JAR, PLDI, and CAV. Dr. Jha has been a Principal Investigator on DoD and US Govt. programs on trustworthy, resilient, and neuro-symbolic AI, including DARPA Assured Autonomy, DARPA Neuro-symbolic Learning and Reasoning, DARPA Symbiotic Design of Cyber-physical Systems, IARPA TrojAI, NSF Self-Improving Cyber-Physical Systems, and Army Research Laboratory’s Internet of Battlefield Things REIGN CRA.