AI-Guided Human-In-the-Loop Inverse Design of High Performance Engineering Structures

TL;DR

This work introduces HiTopAI, an AI co-pilot integrated into human-in-the-loop topology optimization to predict regions of interest for design modification. The AI formulates the prediction as an image-segmentation task using a U-Net trained on synthetically generated masks corresponding to the longest topological member or the most complex node, derived from a skeletonized graph representation of TO designs. Results show that the AI-assisted workflow can achieve substantial performance gains (e.g., a 39% increase in buckling load for an L-bracket) with only a modest overhead in design time, while exhibiting strong generalization to diverse, non-standard topologies and even emergent multi-region predictions. The approach preserves human decision-making, does not alter optimization objectives, and serves as a foundation for future multi-criteria co-pilots and user studies to quantify workflow improvements in industrial design tasks.

Abstract

Inverse design tools such as Topology Optimization (TO) can achieve new levels of improvement for high-performance engineered structures. However, widespread use is hindered by high computational times and a black-box nature that inhibits user interaction. Human-in-the-loop TO approaches are emerging that integrate human intuition into the design generation process. However, these rely on the time-consuming bottleneck of iterative region selection for design modifications. To reduce the number of iterative trials, this contribution presents an AI co-pilot that uses machine learning to predict the user's preferred regions. The prediction model is configured as an image segmentation task with a U-Net architecture. It is trained on synthetic datasets where human preferences either identify the longest topological member or the most complex structural connection. The model successfully predicts plausible regions for modification and presents them to the user as AI recommendations. The human preference model demonstrates generalization across diverse and non-standard TO problems and exhibits emergent behavior outside the single-region selection training data. Demonstration examples show that the new human-in-the-loop TO approach that integrates the AI co-pilot can improve manufacturability or improve the linear buckling load by 39% while only increasing the total design time by 15 sec compared to conventional simplistic TO.

Figures (10)

• Figure 1: (a) Classic TO where the human designer initiates the design process and a fully automated algorithm solves the design problem which the designer judges the final quality of. In contrast, (b) Human Informed Topology Optimization (HiTop) (b.i) allows TO to generate the design solution in collaboration with the human user. In initially generates a quick automated design, and in (b.ii,b.iii), the human designer interactively updates the local minimum feature size requirements to fit their preferences, giving the new design in (b.iv). (c) Illustrates how the integration of an AI-co pilot will improve the design time for HiTop. In (c.i) the human uses multiple trials to identify the topological feature to modify, where as (c.ii) shows how a AI co-pilot recommends where modifications should be made based on previous interaction data.
• Figure 2: Effect on buckling performance of using HiTopAI. The classic TO L-bracket design problem is shown in (a) which here is solved on a $150\times150$ unit$^2$ mesh, and (b) gives the initial topology after 50 iterations. The linear buckling performance for the initial design is $P_{cr}= 6.62\cdot10^{-3}P$ and the buckled shape is shown in (c). The post-processed ML prediction for the human preference region is given in (d), and recommended through the AI co-pilot as the cyan ellipse in (e). In (e), it is also illustrated how the user modifies the recommendation slightly to the purple ellipse. The user updates the minimum feature size within this region as shown in (f), which produces the final HiTopAI design in (g). The buckling performance of the final design is $P_{cr}= 9.20\cdot10^{-3}P$ and the buckled shape shown in (g).
• Figure 3: Example using HiTopAI to modify the most complex node. (a) gives the design problem that is solved on a $150\times 100$ unit$^2$ mesh. (b) shows the initial TO result after 50 iterations. The ML model's output is shown in (c), which displays the post-processed segmentation mask. The co-pilot's suggested intervention is represented by the cyan ellipse in (d), which the human user then fine-tunes into the purple ellipse. (e) illustrates the resulting minimum feature size field updated by the human intervention, leading to the final optimized design with reduced node complexity shown in (f).
• Figure 4: Performance of the human preferences ML model on unseen TO designs from the test dataset. (a) gives examples of the model's performance on 10 randomly designs. The input topologies are given in (a.i), (a.ii) shows the ML model's output (post-processed), and (a.iii) provides the corresponding ground truths. In (b), examples of IOU metric trends observed for the trained ML model of human preference predictions is shown. The first case (b.i) sees the most desirable IOU performance with IOU values of 0.97 which occurs for topologies with a clear longest member. The second case (b.ii), with median IOU values of 0.38, shows that the model's thresholded predictions do not fully match, but tends to overestimate compared to the ground truth. Finally, the third case (b.iii), with IOU values of close to 0, indicates that the synthetically generated dataset is imperfect.
• Figure 5: Examples of the performance of the trained ML model's predictions on TO designs outside the training dataset. Here the model is trained on segmenting the most complex node. Three TO problems are show in (a, e, i), and the corresponding initial TO designs after 50 iterations are given in (b, f, j). The post-processed ML modeling results for human preference for interventions are show as segmentation masks in (c, g, k). For comparison, skeletonized representations of the initial TO designs are given in (d, h, l), with the most complex node highlighted by a dashed line circle.