Our research develops computational methods that enable machines to interpret, generate, analyze, and improve engineered designs. By connecting geometry, physics, data, and human intent, we seek to make engineering design and manufacturing more intelligent, accessible, and responsive to real-world constraints.
This goal has guided VDEL from its early work in sketch-based modeling and learned shape representations to current research in generative design, scientific machine learning, intelligent manufacturing, and engineering-focused vision-language systems. Across these areas, we are interested not only in predicting an answer, but also in representing the engineering information, constraints, and physical relationships needed to produce useful and reliable results.
Our work is organized around four connected research directions.
Engineering knowledge appears in many forms: sketches, drawings, diagrams, CAD models, meshes, physical fields, requirements, and systems models. For a computational system to support design, it must first represent and interpret this information in a form that preserves engineering meaning.
VDEL’s early research developed pen- and sketch-based methods for constructing and editing three-dimensional geometry. We subsequently introduced methods for jointly analyzing collections of shapes, learning common deformation structures, and controlling geometry through semantic concepts rather than conventional CAD operations. This work helped establish a progression from direct geometric interaction to learned representations of design intent.
Today, we are extending this research to engineering documents and structured models. Current projects investigate graph-based understanding of technical drawings, generation of editable CAD from orthographic views, and translation of natural-language requirements into SysML models. The long-term objective is to connect informal human communication with the structured representations required for engineering analysis and downstream redesign.
Representative work
Generative engineering requires more than producing plausible geometry. A viable design must satisfy physical requirements, manufacturing constraints, cost objectives, and the realities of available processes and resources.
Our work in this area combines geometry, optimization, machine learning, and engineering analysis. We have developed methods for designing lightweight structures under uncertain loading, coupling structural and process decisions in metal additive manufacturing, synthesizing three-dimensional forms from design data, and accelerating topology optimization using learned physical fields.
More recently, we have introduced generative manufacturing, a framework in which products are redesigned in response to both functional requirements and available manufacturing resources. This direction expands generative design beyond shape creation toward the coordinated reasoning needed when materials, machines, processes, suppliers, or mission requirements change.
Representative work
High-fidelity engineering simulation can be computationally expensive, particularly when it must be repeated across many geometries, operating conditions, or design iterations. We investigate how machine learning can accelerate these calculations without discarding the geometry and physics that govern them.
Our research has developed neural methods for stress and scalar-field prediction, graph- and mesh-based learning, topology generation, and reduced-order modeling. These approaches operate on engineering representations rather than only regular images, allowing predictions to respect complex domains and discretizations. Current work explores neural fields, operator-learning architectures, low-rank attention, and geometry-conditioned models for large-scale physical systems.
The goal is not to replace physical reasoning with a black-box approximation. We seek computational models that combine the speed and adaptability of learning with the structure, constraints, and interpretability required in engineering analysis.
Representative work
Manufacturing processes generate rich streams of acoustic, thermal, visual, geometric, and process data. We develop methods that transform these measurements into useful estimates of process state, product quality, and manufacturing risk.
Our work in metal additive manufacturing includes acoustic and thermal monitoring, inference of melt-pool behavior, prediction of lack-of-fusion defects, and analysis of fast physical phenomena during laser-material interaction. These projects combine sensing, machine learning, mechanics, and manufacturing science to reveal events that may be difficult or impossible to observe directly during production.
We also investigate manufacturability prediction and resource-aware redesign. Together, these efforts support a broader vision in which engineering systems can interpret available evidence, anticipate production problems, and adapt designs to changing manufacturing conditions.
Representative work
The computational ideas developed in VDEL have been tested across a range of engineering and scientific domains. Applications have included additive manufacturing, integrated-circuit design, fuel cells, environmental sensing, molecular and peptide design, patient-specific modeling, soft-tissue deformation, and organ preservation.
These collaborations are not separate from the laboratory’s central research identity. They provide demanding settings in which representations, learning methods, physical models, and design algorithms can be evaluated against consequential real-world problems.
Examples include generative design of target-specific peptide inhibitors in Nature Communications and a computational liver perfusion model in Scientific Reports.
VDEL brings together students and collaborators with backgrounds in mechanical engineering, computer science, applied mathematics, manufacturing, and related fields. Our projects typically combine methodological research with computational implementation and validation on substantive engineering problems.