Inverse Design of Materials Across the Process–Structure–Property Linkage
Inverse design has become a key paradigm in modern materials engineering, promising to shift material development from trial-and-error exploration toward goal-driven design. Instead of asking how a given process or microstructure performs, inverse design starts from desired properties and seeks to identify material structures and processing paths that can realize them.
A central difficulty arises from the tightly coupled nature of the process–structure–property relationship. Material properties are not directly controllable design variables; they emerge from microstructural features that themselves result from complex, history-dependent manufacturing processes. As a consequence, inverse design problems are typically ill-posed: many different microstructures may exhibit similar properties, while not all of them are physically realizable or reachable by a given process. Classical optimization approaches often ignore this ambiguity and implicitly assume a one-to-one mapping between properties and structures, leading to solutions that are optimal in theory but infeasible in practice.
This challenge is particularly pronounced in manufacturing scenarios where process constraints play a ominant role. Even if a microstructure with ideal properties can be identified computationally, the corresponding processing path may be inaccessible due to physical limitations, stability issues, or economic constraints. From an inverse design perspective, the key question therefore shifts from identifying the optimal microstructure to identifying which of many acceptable solutions can actually be produced.
Our research addresses inverse design from this broader, manufacturing-aware perspective. Instead of collapsing the inverse problem to a single target solution, it explicitly treats inverse design as a search for sets of equivalent solutions. Desired properties define a target region rather than a single point, allowing multiple microstructures to be considered acceptable from a performance standpoint. This reframing transforms non-uniqueness from a limitation into an advantage: by generating multiple candidate solutions, inverse design can be coupled with process considerations to select the most reachable outcome.
Within this context, machine learning serves as an enabling technology. Learned representations areused to capture complex structural information, making it possible to explore inverse mappings thatwould be intractable with direct simulation alone. Demonstrated on a metal forming scenario wherecrystallographic texture governs elastic and anisotropic properties, the approach illustrates how inversedesign can move beyond idealized solutions toward manufacturable ones. More generally, itexemplifies a shift in inverse design thinking: from identifying optimal structures in isolation to navigating feasible regions of the design space that respect both property objectives and process realities.
Related Publications:
Contact:
Prof. Dr. Stefan Sandfeld
Tel.: +49 241/927803-11
E-mail: s.sandfeld@fz-juelich.de