Abstract
Additive manufacturing has become one of the forefront technologies in fabrication, enabling new products impossible to manufacture before. Although many materials exist for additive manufacturing, they typically suffer from performance trade-offs preventing them from replacing traditional manufacturing techniques. Current materials are designed with inefficient human-driven intuition-based methods, leaving them short of optimal solutions. We propose a machine learning approach to accelerate the discovery of additive manufacturing materials with optimal trade-offs in mechanical performance. A multi-objective optimization algorithm automatically guides the experimental design by proposing how to mix primary formulations to create better-performing materials. The algorithm is coupled with a semi-autonomous fabrication platform to significantly reduce the number of performed experiments and overall time to solution. Without any prior knowledge of the primary formulations, the proposed methodology autonomously uncovers twelve optimal composite formulations and enlarges the discovered performance space 288 times after only 30 experimental iterations. This methodology could easily be generalized to other material formulation problems and enable completely automated discovery of a wide variety of material designs.
Introduction
In additive manufacturing (AM), innovation is frequently bottlenecked by material limitations rather than hardware capabilities. Engineering high-performance materials involves navigating complex trade-offs; for instance, increasing a material's stiffness (modulus) often results in a corresponding loss of toughness. Traditional discovery methods rely on "one factor at a time" (OFAT) testing and human intuition, which are ill-suited for identifying the "Pareto front"—the set of formulations representing the optimal balance between conflicting mechanical properties.
To address this, researchers developed a system that utilizes eight commercially available chemical components—including photoinitiators, diluents, and oligomers—to create a set of primary formulations. As shown in the initial performance data, these base materials exhibit widely varying stress-strain behaviors, providing a diverse but unoptimized starting point for discovery.
Methods: The Closed-Loop Pipeline
The core of this research is a semi-automated, closed-loop hardware and software pipeline designed for rapid iteration. The process begins with Automated Dispensing & Mixing, where a 6-channel syringe pump system dispenses primary formulations into a dual asymmetric centrifuge mixer. Samples are then fabricated using a Jet-Valve 3D Printer, chosen for its ability to handle a broad range of viscosities without the frequent recalibration required by standard extrusion or inkjet systems.
The "brain" of the system is a Multi-Objective Bayesian Optimization algorithm. It employs Gaussian Processes (GP) as surrogate models to predict the performance of untested mixtures. To balance the exploration of new chemical spaces with the exploitation of known high-performing regions, the system uses Thompson Sampling. Each iteration, the algorithm proposes a batch of four new formulations to be tested in parallel, significantly accelerating the search process.
Results and Discussion
The data-driven approach yielded significant improvements over traditional methods. Over 30 iterations, the system evaluated 150 unique formulations (starting from an initial seed of 30).
Key achievements include:
- Discovery of 12 Pareto Optimal Materials: The algorithm successfully mapped the optimal trade-offs within a 6-dimensional design space.
- Performance Gains: The optimized materials showed a 70.8% increase in maximum compression strength and a 50.8% increase in toughness compared to the best-performing initial primary inks.
- Efficiency: The "performance space" (hypervolume) of discovered materials expanded by a factor of 288 compared to the starting points.
- Chemical Insights: The optimization identified Urethane Dimethacrylate (UDMA) as a critical component for high stiffness, while specific ratios of urethane-modified acrylates were found to be essential for maximizing toughness.
Conclusion
This methodology demonstrates that coupling machine learning with semi-autonomous fabrication can drastically reduce the time and experimental overhead required for material discovery. While currently "semi-automated"—requiring human assistance for sample transfer—future iterations could achieve full autonomy through robotic integration. This workflow is highly generalizable and could be applied to the discovery of various functional materials, from surgical sealants to specialized industrial coatings.
Summary updated