(135c) Paraflow: Generative Design for Faster Additive Fabrication with Fewer Supports | AIChE

(135c) Paraflow: Generative Design for Faster Additive Fabrication with Fewer Supports


Lipkowitz, G. - Presenter, Stanford University
DeSimone, J. M., University of North Carolina at Chapel Hill
Shaqfeh, E., Stanford University
Coates, I., Stanford
Krishna, N., Stanford
While 3D printing affords designers unprecedented geometric complexity in their products, currently the technology requires cumbersome support structures that are not user-friendly in numerous ways: materially wasteful, human labor-intensive, time-consuming to remove, damaging to surface finish, and often unreliable in ensuring printability in the first place. Inspired by recent physical demonstrations that injecting liquid through an embedded channel can neutralize suction during vat 3D printing, we discover this phenomenon can also significantly reduce the quantity of supports required by a user's model, or entirely obviate them for certain product geometries. To allow any resin 3D printer user to take advantage of this approach, we develop a novel fluid dynamics-guided computational inverse design tool, Paraflow, which innervates the user's arbitrary to-be-printed 3D model with a monotonic fluidic network for injection during printing. Our work draws on recent progress in generative artificial intelligence, in particular applied to search algorithms for 3D data trained on open source computer aided design (CAD) repositories such as Shapenet and Thingi10k, to optimize this network. Our goal is, for an arbitrary 3D model provided as input by a CAD modeler and 3D printer user, to construct a fluidic network that offsets suction, in order to effectively replace, or minimize, support structures required.

To generate such a suitable 3D printable network, we formulate the design problem as a path planning optimization problem where, for an arbitrary 3D geometry to be 3D printed via a set of 2D slice layers, we computationally design a corresponding fluidic network that innervates the part to sufficiently distribute one, or multiple, materials during printing. Our inverse design approach originates a fluidic network with a single input injection point connected to the build platform and, thus, the syringe pump. Then, injection channels are positioned within the user-provided model at optimal positions as determined by surrogate fluid dynamics modeling, in order to offset suction during printing to minimize need for supports. As it grows in complexity, the network is modeled as an electrical circuit, as is typical for pressure-driven fluidics. Branches may be either open (with outlets), or closed (without outlets), thus not contributing to fluid flow. The final network, encoded as a graph connectivity matrix of nodes and edges, is converted to a series of smooth NURBS curves, swept to produce a positive network 3D geometry, and finally Boolean differenced with the original B-rep CAD model via surface-to-surface intersection to produce negative channels, hence an innervated part. Rather than presenting the user with a single generatively designed fluidic network, Paraflow presents multiple options for innervating networks to the user, some of which may be more desirable for functional and/or aesthetic requirements. While there is one optimal network configuration that minimizes suction during printing, many possible configurations satisfy the design space constraints described above, with higher variance generally in part footprints with higher cross sectional area.

We experimentally demonstrate that Paraflow, applied to an off-the-shelf 3D printer with an inexpensive syringe pump add-on extension, enables designers to fabricate otherwise unprintable overhang geometries without support structures. When supports are not added to the model in the traditional process, print defects ranging from minor but noticeable to catastrophic result. With injection 3D printing and Paraflow computational design, however, such bridge overhangs are readily printable unsupported. To validate our fluidic inverse design approach, we describe our set-up for measuring forces during resin printing. In brief, we position both a load cell on the platform of our printer, to record suction forces, and a camera underneath the vat, to visualize the flow of injected resin. These online force sensors demonstrate that flow through these generatively-designed networks significantly reduce forces during printing, compared with control experiments without injection. This effect scales with the number of injection sites, as expected, and with the area of the printed object.

Finally, elastomers are among the most challenging materials with which to vat 3D print, due to aforementioned suction forces, but also among the most attractive for their energy absorbing or returning properties. We demonstrate that by overcoming suction with our generative design tool, we can print with highly flexible, shape-changing materials for applications in soft robotics and deformable interfaces, without support structures. Importantly, we characterize the limitations of our injection 3D printing generative design strategy, including the necessity to integrate hollow channels of non-zero thickness into parts, and describe post-curing strategies for mitigating this limitation. We also outline future directions for our novel generative AI design tool, including developing further software for fluid injection through multimaterial lattices, and integrating real-time computer vision control via neural networks trained on synthetically-generated data into injection 3D printers to allow users to better control multimaterial flows in real-time.