Knowledge-informed Bayesian optimization resolves multi-objective trade-offs in 3D printing photopolymers

Technical Summary: Knowledge-Informed Bayesian Optimization for 3D Printing Photopolymers

Problem Statement
The discovery of high-performance materials is frequently hindered by high-dimensional design spaces, competing objectives, and limited experimental budgets. This challenge is particularly acute in the development of 3D printing photopolymers, where advanced engineering applications demand the simultaneous optimization of mechanical strength, thermal resistance, toughness, and printability. Conventional trial-and-error approaches struggle with the coupled trade-offs inherent in these systems (e.g., network rigidity vs. segmental mobility). Furthermore, existing data-driven methods, such as standard Bayesian optimization, often rely on static search domains or predefined constraints. When feasible boundaries and high-performance regions are only revealed through iterative experimentation, static formulations can lead to inefficient sampling, premature convergence, or an inability to access target regions effectively.

Methodology: KI-DC-BO Framework
The authors propose a Knowledge-Informed Dynamic-Constrained Bayesian Optimization (KI-DC-BO) framework. Unlike conventional approaches that primarily update surrogate models while keeping constraints static, KI-DC-BO integrates experimental feedback to dynamically refine both the surrogate model and the feasible domain.

The workflow operates as a closed loop:

1. Surrogate Modeling: At each iteration $t$, accumulated experimental data ($\mathcal{D}_t$) trains Gaussian Process (GP) surrogate models for the objectives.
2. Candidate Generation: The q-NEHVI (expected hypervolume improvement) acquisition function proposes candidate formulations within the current feasible domain ($\Omega_t$).
3. Dynamic Constraint Refinement: This is the core innovation. Domain knowledge derived from experimental results is converted into updated feasibility constraints. These constraints reshape the search space for the next iteration, redirecting sampling toward regions of increasing multi-objective value.
4. Secondary Screening: Candidates are filtered by knowledge-guided actionability criteria (e.g., printability) before material preparation.
5. Feedback: Experimental results update the surrogate model and inform the next round of constraint updates.

The method utilizes "regret-derived diagnostic metrics" (specifically the regret coefficient $\beta_k$) to evaluate convergence and determine when feasible-domain refinement is necessary.

Application Case: 3D Printing Photopolymers
The framework was applied to optimize a photopolymer system based on THEICTA (Tris(2-hydroxyethyl) isocyanurate triacrylate), a high-$T_g$ base component, blended with various monomers:

• Monofunctional regulators: IBOA and ACMO (to modulate segmental motion and stress relaxation).
• Multifunctional modifiers: DCPDA, NPGDA, and DITMPTA (to tune crosslinking and thermomechanical response).

The optimization targeted three competing objectives:

• Tensile Strength ($Y_1$)
• Elongation at Break ($Y_2$)
• Glass-Transition Temperature ($Y_3$)

Key Contributions and Results
Through 15 iterative rounds and 90 experiments, KI-DC-BO successfully identified multiple formulations exhibiting balanced multi-objective performance.

• Dynamic Constraint Evolution:
  • Initial State: Constraints were set based on monomer-blending priors (e.g., $X_2 \ge 10$ for ACMO).
  • Update I: Feature samples indicated that higher ACMO content improved the strength-elongation balance. The constraint was refined to $X_2 \ge 20$.
  • Update II: Further analysis showed that high ACMO content required limited multifunctional monomer loading to stabilize the balance. Constraints were refined to $X_2 \ge 30$ and $X_3 + X_4 + X_5 \le 10$.
• Performance Metrics:
  • The "high-performance hit rate" (formulations meeting $Y_1 \ge 90$ MPa, $Y_2 \ge 2.0\%$, $Y_3 \ge 210^\circ$C) increased from 2/30 in initial samples to 15/20 after dynamic constraint refinement.
  • Representative formulations achieved tensile strengths >100 MPa, glass-transition temperatures >210 °C, and elongations at break of 2–3%.
  • These formulations retained the ability to print complex, high-resolution lattice structures (feature size ~0.1 mm).
• Comparison: The optimized photopolymers outperformed current commercial 3D printing photopolymers (e.g., Carbon EPX 86FR, LOCTITE 3D IND147) in both heat resistance and tensile strength, achieving a superior balance that commercial systems typically struggle to maintain.

Significance and Claims
The paper claims that KI-DC-BO establishes dynamic feasible-domain refinement as a practical strategy for materials discovery under high-dimensional, constrained, and multi-objective conditions. By converting experimentally derived domain knowledge into iteratively updated feasibility constraints, the framework allows the search space to restructure itself, redirecting limited experimental budgets toward regions of synergistic multi-objective performance.

The authors emphasize that this approach resolves coupled trade-offs (strength, thermal resistance, elongation, printability) more efficiently than static constraint methods. While demonstrated on 3D printing photopolymers, the framework is presented as a transferable strategy for any materials discovery problem where feasible boundaries are unknown a priori and must be revealed through iterative experimentation.