Public Abstract

DE-SC0024852: AI Enabled Co-Design of Polymer AM Targets

Award Status: Expired

Institution: UHV3D Inc, Santa Fe, NM
UEI: GKA8EBE195L3
DUNS:

Most Recent Award Date: 12/10/2024
Number of Support Periods: 1
PM: Akli, Kramer

Current Budget Period: 02/12/2024 - 01/31/2025
Current Project Period: 02/12/2024 - 01/31/2025
PI: Perumal, Vignesh

Supplement Budget Period: N/A

Public Abstract

AI Enabled Co-Design of Polymer AM Targets-UHV3D Inc, 2778 Agua Fria Street, STE 4, Santa Fe, NM 87507 Vignesh Perumal, Principal Investigator, vignesh@uhv3d.com Simon Woodruff, Business Official, simon@uhv3d.com Amount: $199,999 A principal cost driver in the IFE system is the target. Current designs will not be viable in a fusion energy system, and there is considerable effort to design targets that will be manufactured at scale with the requisite uniformity. Recent advances with wetted-foam targets show that they naturally develop a shell which makes them suitable for mass production [1]. These offer several advantages, including simplicity in target production (suitable for mass production for inertial fusion energy), absence of the fill tube (leading to a more-symmetric implosion), and lower sensitivity to both laser imprint and physics uncertainty in shock interaction. There are now many methods for making foam targets, although one stands out as not only scalable but also optimizable - 2 photon polymerization - developed at GA by Haid et al [2]. The desire to put nuclear fusion on the grid will place some of the most extreme demands on materials. Future milestones for fusion will depend on, among other things, the ability to push the limits of mechanical design to extract maximum performance by increasing the gain and shell stability. The outstanding issues of efficiency during performance can be addressed by considering the geometric parameter optimization of the wetted foam with respect to the operating conditions and laser interaction. For instance, the thickness of the foam can be optimized such that the foam layer is completely ablated while addressing the hydrody- namic effects such as Rayleigh-Taylor instability. This requires multiple different physics-based simulations that can then be fed into an optimization algorithm. However, this design paradigm relies on developing complex multiscale modeling approaches to quantify how coupled physical processes at the nanometer (nm) length scale drive the performance of a part in the meter length scale. In practice, this paradigm is typically plagued with two main limitations: (i) transition across length scales necessitates downsampling information, thus potentially filter useful information; (ii) this computational framework is computationally expensive, practically infeasible for multiple design development runs, and rather complex to use. ML/AI offers a route to change this paradigm by generating surrogate models that emulate the response surface of more computationally expensive models [3]. These surrogates can capture scale transitions in a computationally inexpensive framework compared to hierarchical scale transition approaches using high-fidelity simulations. Further surrogates provide a means to rapidly calibrate models against exper- imental data (i.e. formal model calibration techniques), to quantify model form uncertainty, and to quantify parametric uncertainty. We propose an agile approach to engineering design software development using data-driven techniques. An optimized wetted-foam design that addresses the above problem will be the minimum viable product. We show the proposed AI/ML for developing multidisciplinary design optimization approach in Fig. 2. This approach leverages the use of FEM software to develop a database. Thereafter, AI/ML models such as deep neural networks (DNNs) will be used to learn the underlying physics from the database. The simulations are intended to capture multiple physical responses from structural, fluid, and radiation physics. These physical responses account for geometric performance during service and also the build stresses during the manufacturing of the part. The database is used to train features of the DNNs. DNNs are known to efficiently learn and capture the underlying complex relationships between the input parameters, such as in- put current, material, topology, and the resulting physical responses (stress, displacement, internal pressure) of the part under such inputs [4]. The objective and constraints of the optimization are also defined based on the requirements of the design. The DNN models will then act as a relatively inexpensive surrogate to the more expensive FEM simulations and compute the various physical quantities and their gradients that will be fed into a gradient-based topology optimization. Using the DNNs in place of the standard FEM as solvers can drastically reduce the computational time for iterative approaches such as optimization while ensuring the desired accuracy [5]. This optimization loop would provide an optimal design. This proposed workflow can be verified using standard FEM models in the design loop. In Phase I, we will develop the modeling capability and exercise the workflows using open-source computational tools aiming to optimize the design point for wetted-foam targets. In Phase II, we will be seeking to print the optimized build designs in concert with collaborators and validate the design by manufacturing and testing it in real-world conditions. There is currently a demand for the above-mentioned capability in fusion energy development, particularly by the private sector. Demonstrating that we can design optimized wetted-foam targets and predict performance will mean that the tools will be adopted. We will consider a licensing model in Phase II, and develop the currently open-source capabilities further offering those under a standard open source license or developing proprietary applications. Commercialization will therefore usual SaaS methods - we have done this successfully with SciVista, previously supported by a SBIR grant. All of our SBIR grants have in fact led to commercial success - all 4 Phase I’s have led to Phase II’s and successful businesses as a consequence.

AI Enabled Co-Design of Polymer AM Targets-UHV3D Inc, 2778 Agua Fria Street, STE 4, Santa Fe, NM 87507

Vignesh Perumal, Principal Investigator, vignesh@uhv3d.com

Simon Woodruff, Business Official, simon@uhv3d.com

Amount: $199,999

A principal cost driver in the IFE system is the target. Current designs will not be viable in a fusion energy system, and there is considerable effort to design targets that will be manufactured at scale with the requisite uniformity. Recent advances with wetted-foam targets show that they naturally develop a shell which makes them suitable for mass production [1]. These offer several advantages, including simplicity in target production (suitable for mass production for inertial fusion energy), absence of the fill tube (leading to a more-symmetric implosion), and lower sensitivity to both laser imprint and physics uncertainty in shock interaction. There are now many methods for making foam targets, although one stands out as not only scalable but also optimizable - 2 photon polymerization - developed at GA by Haid et al [2]. The desire to put nuclear fusion on the grid will place some of the most extreme demands on materials. Future milestones for fusion will depend on, among other things, the ability to push the limits of mechanical design to extract maximum performance by increasing the gain and shell stability. The outstanding issues of efficiency during performance can be addressed by considering the geometric parameter optimization of the wetted foam with respect to the operating conditions and laser interaction. For instance, the thickness of the foam can be optimized such that the foam layer is completely ablated while addressing the hydrody- namic effects such as Rayleigh-Taylor instability. This requires multiple different physics-based simulations that can then be fed into an optimization algorithm. However, this design paradigm relies on developing complex multiscale modeling approaches to quantify how coupled physical processes at the nanometer (nm) length scale drive the performance of a part in the meter length scale. In practice, this paradigm is typically plagued with two main limitations: (i) transition across length scales necessitates downsampling information, thus potentially filter useful information; (ii) this computational framework is computationally expensive, practically infeasible for multiple design development runs, and rather complex to use. ML/AI offers a route to change this paradigm by generating surrogate models that emulate the response surface of more computationally expensive models [3]. These surrogates can capture scale transitions in a computationally inexpensive framework compared to hierarchical scale transition approaches using high-fidelity simulations. Further surrogates provide a means to rapidly calibrate models against exper- imental data (i.e. formal model calibration techniques), to quantify model form uncertainty, and to quantify parametric uncertainty. We propose an agile approach to engineering design software development using data-driven techniques. An optimized wetted-foam design that addresses the above problem will be the minimum viable product. We show the proposed AI/ML for developing multidisciplinary design optimization approach in

Fig. 2. This approach leverages the use of FEM software to develop a database. Thereafter, AI/ML models such as deep neural networks (DNNs) will be used to learn the underlying physics from the database. The simulations are intended to capture multiple physical responses from structural, fluid, and radiation physics. These physical responses account for geometric performance during service and also the build stresses during the manufacturing of the part. The database is used to train features of the DNNs. DNNs are known to efficiently learn and capture the underlying complex relationships between the input parameters, such as in- put current, material, topology, and the resulting physical responses (stress, displacement, internal pressure) of the part under such inputs [4]. The objective and constraints of the optimization are also defined based on the requirements of the design. The DNN models will then act as a relatively inexpensive surrogate to the more expensive FEM simulations and compute the various physical quantities and their gradients that will be fed into a gradient-based topology optimization. Using the DNNs in place of the standard FEM as solvers can drastically reduce the computational time for iterative approaches such as optimization while ensuring the desired accuracy [5]. This optimization loop would provide an optimal design. This proposed workflow can be verified using standard FEM models in the design loop. In Phase I, we will develop the modeling capability and exercise the workflows using open-source computational tools aiming to optimize the design point for wetted-foam targets. In Phase II, we will be seeking to print the optimized build designs in concert with collaborators and validate the design by manufacturing and testing it in real-world conditions. There is currently a demand for the above-mentioned capability in fusion energy development, particularly by the private sector. Demonstrating that we can design optimized wetted-foam targets and predict performance will mean that the tools will be adopted. We will consider a licensing model in Phase II, and develop the currently open-source capabilities further offering those under a standard open source license or developing proprietary applications. Commercialization will therefore usual SaaS methods - we have done this successfully with SciVista, previously supported by a SBIR grant. All of our SBIR grants have in fact led to commercial success - all 4 Phase I’s have led to Phase II’s and successful businesses as a consequence.