Space Structures Optimization Using AI Algorithms

Tekrevolution is performing structural optimization (space domain components) with the aim both to pursue metal replacement and to use techno-polymers Additive Manufacturing (FDM).?

To do that we are using surrogate models driven by Artificial Intelligence coupled to commercial topology optimization codes.?

The component shown is an FDM European Space Agency - ESA satellite antenna bracket (Carbon PEEK).

AI algorithms we use belong to our partner UptimAI and, for this problem, we used two brand new developed optimization features.?

The first one is the Direct Optimizer which aims at minimizing (or maximizing) a function of interest with one single output with so-called Evolutionary Algorithms which mimic the behavior of a population of living beings in an environment subject to natural selection. Inside the Direct Optimizer we are proposing two different algorithms, depending on the problem: Differential Evolution (DE).?

This algorithm creates a population of potential samples and makes it evolve at each generation. The evolution process involves independent and random mutation of each coordinate of each member of the population. This evolution process is repeated multiple times in a looping fashion and at each iteration, the population members are replaced by the new members (children) only if they provide a better result (lower function value).?

The algorithm stops when the population has converged closely enough to a minimum point or the maximum allowed number of function evaluations is attained. Because each coordinate is mutated independently this algorithm is subject to the curse of dimensionality. This is why the decoupling option is recommended for high dimensional problems (dimension >= 10).

The second one is the Hybrid method of DE and Covariance Matrix Adaptation. The Covariance Matrix Adaptation (CMA) is an algorithm that evolves the population by generating children according to a multivariate Gaussian distribution. The best child elements are selected for the next population and the parameters of the Gaussian distribution are adjusted in a heuristic manner. Again the algorithm stops in case of convergence of the population or if the maximum number of function evaluations has been reached. The Hybrid Method performs a decoupling process to separate 1D-subproblems and ND-subproblems. Then, the DE algorithm is applied to 1D-subproblems and the CMA algorithm is applied to ND-subproblems.

To tackle the curse of dimensionality, for both algorithms, the software use a decoupling process (also known as variable separation in the literature) that transforms the main domain into smaller subdomains. Then the optimization process is separated into lower-dimensional subproblems, providing better performance and converging faster.

On the other hand, the Multi-Objective Optimizer is designed to minimize multiple concurrent outputs.?

It works in a similar way that the Direct Optimizer does.?

However, it is not usable with decoupling and uses an evolutionary algorithm called NSGA-II.?

Unlike the Direct Optimizer, the Multi-Objective Optimizer provides multiple solutions approximating a Pareto Front of the problem.

Optimization works for small components as well as for large space infrastructures?aiming at reducing scrap materials, weight and complexity allowing use of new materials and new manufacturing processes. This will saving time and costs and will be one of the pillars on which new space programs will rely for a further technological evolution. Also many other industry sectors will benefit from this capabilities to improve their products design, performance and manufacturing.