Range-Aware Bayesian Optimization for Discovering Diverse Designs within Target Property Windows

This paper introduces a range-aware Bayesian optimization framework that efficiently discovers diverse designs satisfying target property ranges by directly scoring the posterior probability of range compliance, demonstrating superior performance over standard methods in both benchmarks and practical materials design case studies.

The Gist

Imagine you are a chef trying to invent a new soup. Most traditional cooking competitions ask you to find the single best soup possible—the one with the absolute highest flavor score. You might spend all your time tweaking that one recipe until it's perfect.

But in the real world, especially when designing new materials or products, you often don't need the "perfect" soup. You just need a soup that tastes good enough. It needs to be salty enough, but not too salty; hot enough, but not scalding. You have a "target range" of acceptable flavors. Furthermore, you don't just want one good soup; you want a menu of different options. Maybe one is cheaper to make, another is easier to cook, and a third uses ingredients you already have.

This paper introduces a new "smart cooking assistant" (a mathematical tool called Range-Aware Bayesian Optimization) designed specifically to find that menu of good-enough options, rather than just hunting for the single perfect one.

The Problem with the Old Way

Traditional "smart assistants" (standard optimization methods) are like chefs obsessed with perfection. They look at a recipe and ask, "Is this better than the best one I've seen so far?" If the answer is yes, they keep going. If they find a soup that is already "good enough," they might stop looking for other options and just keep tweaking that one bowl to make it slightly better.

This is a problem because:

1. They miss the variety: They might find one great soup but ignore ten other soups that are also perfectly good but taste slightly different.
2. They get stuck: They might focus all their energy on one tiny corner of the kitchen, missing other areas where great soups could be hiding.

The New Solution: The "Range-Aware" Assistant

The authors, Shengli Jiang and colleagues from Princeton University, built a new assistant that thinks differently. Instead of asking, "Is this the best?", it asks, "What is the probability that this recipe falls within my acceptable range?"

They call their best method the "Tolerance Ball" (TB).

Here is how it works using an analogy:
Imagine you are throwing darts at a wall.

• The Old Way: You are trying to hit the exact bullseye. If you get close, you keep throwing at that same spot to get closer.
• The New Way (Tolerance Ball): You have a large, fuzzy circle drawn on the wall. You don't care about the bullseye; you just want to hit anywhere inside that circle. The new assistant calculates the odds that your next dart will land inside that circle. If a spot on the wall has a high chance of landing inside the circle, it sends a dart there.

Because it is looking for any hit inside the circle, it naturally spreads its darts out to find different spots within that circle, rather than clustering them all in one spot. This gives you a diverse set of valid recipes.

How They Tested It

The team tested this new assistant in two main ways:

1. The Video Game Level (Benchmarks): They used standard math puzzles where the goal was to find inputs that produced specific outputs. They compared their new "Tolerance Ball" method against old methods (like "Expected Improvement") and random guessing.

   • Result: The new method found more valid solutions and a wider variety of them than any other method. It was like finding 10 different keys that open the same door, while the old methods only found one key or kept trying to polish that one key.

2. Real-World Kitchen Tests (Case Studies):

   • Test 1: Making Plastic (Polymer Synthesis): They tried to find the right cooking conditions (temperature, time, etc.) to make plastic with a specific weight distribution. The goal wasn't just "light" or "heavy" plastic, but a specific shape of the weight curve.
     • Result: The new method found many different combinations of cooking conditions that produced the exact same plastic quality. This is huge for manufacturers because if one method is too expensive, they can switch to a different valid method found by the assistant without changing the product.
   • Test 2: Designing Light-Absorbing Molecules: They looked for specific molecules that absorb light in a certain pattern (useful for things like solar cells or sensors).
     • Result: The assistant found different chemical structures that looked completely different but produced the exact same light-absorption pattern. This gives chemists flexibility to choose the molecule that is easiest or cheapest to build.

Why This Matters

The paper concludes that for many real-world design problems, we don't need a single "perfect" answer. We need a portfolio of good options.

The "Range-Aware" method is like a smart scout that doesn't just look for the highest mountain peak. Instead, it maps out all the flat, habitable plateaus in a specific altitude range. It tells you: "Here are five different places you can build a house that are all safe, comfortable, and within your budget."

By focusing on the probability of being "good enough" rather than "the best," this new tool helps scientists and engineers discover a richer, more diverse set of solutions, saving time and money while offering more flexibility in how they build their products.

Technical Summary

Technical Summary: Range-Aware Bayesian Optimization for Discovering Diverse Designs within Target Property Windows

1. Problem Formulation

In many materials and product design contexts, the objective is not to find a single global optimum but to identify multiple distinct candidates whose properties fall within prescribed "target ranges." Standard Bayesian Optimization (BO) and multi-objective approaches are typically designed to locate extrema or Pareto frontiers, often overlooking valid interior solutions that satisfy specification windows. Furthermore, existing goal-seeking methods (e.g., Level-set estimation, BAX) or constrained BO formulations often prioritize boundary crossings, treat feasibility as a secondary constraint, or incur high computational costs in high-dimensional spaces.

The authors frame the problem as an inverse design task where the goal is to recover a diverse set of valid designs $S = \{x_1, \dots, x_N\}$ such that the property vector $f(x)$ lies within a Euclidean tolerance $\varepsilon$ of a target specification $y_{tgt}$. The challenge is to achieve this under a limited evaluation budget while ensuring the discovered solutions are distinct and correctly aligned with specific target ranges, particularly when pursuing multiple targets in parallel.

2. Methodology

The paper proposes a Range-Aware Bayesian Optimization (BO) framework centered on two novel, sampling-free acquisition functions: Tolerance Ball (TB) and Heaviside (HV).

• Surrogate Modeling: The framework utilizes Gaussian Process (GP) regression with Matérn 5/2 kernels to model the mapping from design space to property space. For multi-dimensional outputs, independent GPs are fitted to each dimension.
• Acquisition Functions:
  • Tolerance Ball (TB): This acquisition function directly scores the posterior probability that a candidate satisfies the target range. It approximates the predictive distribution of the property vector as isotropic and calculates the probability that the squared discrepancy $d(x) = \|f(x) - y_{tgt}\|^2$ falls below $\varepsilon^2$. This is computed analytically using the cumulative distribution function of a non-central $\chi^2$ distribution.
  • Heaviside (HV): This function prioritizes candidates whose posterior mean lies inside the tolerance range but retains a smooth transition near the boundary. It blends a binary indicator (based on the posterior mean) with the TB probability, controlled by a smoothing parameter $\tau$.
• Parallel Search: The framework extends naturally to parallel search across multiple target ranges ($T$ targets). In this mode, a shared GP surrogate models all targets. At each iteration, each target proposes a candidate by maximizing its specific acquisition function. Duplicates are removed to ensure efficient batch evaluation.
• Baselines: The proposed methods are compared against standard BO acquisition functions (Expected Improvement, Lower Confidence Bound), goal-seeking methods (Posterior-mean BAX), and random sampling.

3. Key Contributions

• Novel Acquisition Functions: The introduction of TB and HV, which treat range satisfaction as the primary acquisition objective rather than a constraint or a byproduct of extremum seeking.
• Analytical Efficiency: The TB acquisition is derived in closed form, avoiding the need for expensive sampling-based set estimation (unlike BAX or MultiBAX), making it computationally efficient for continuous and discrete domains.
• Parallel Target Handling: A mechanism for simultaneously pursuing multiple distinct specifications within a shared design space without compromising target fidelity.
• Diversity Metrics: The use of $\delta$-uniqueness and normalized diversity scores to evaluate the ability of algorithms to recover distinct valid solutions rather than clustering around a single point.

4. Results

The framework was evaluated across four classes of tasks: synthetic benchmark functions, pool-based datasets (nanoparticles, proteins, molecular libraries), kinetic Monte Carlo (KMC) polymerization simulations, and sequence-defined oligomer discovery.

• Benchmark Performance: Across synthetic and pool-based datasets, the Tolerance Ball (TB) acquisition consistently achieved the highest average rank and the largest fraction of best-performing tasks. It outperformed standard BO (EI, LCB) and goal-seeking baselines (BAX) in recovering diverse valid designs. TB maintained balanced discovery across multiple targets, whereas methods like EI tended to concentrate on a subset of targets or refine a single incumbent.
• Target Fidelity: In parallel search scenarios, TB demonstrated superior "target fidelity," exhibiting low cross-target hit ratios (i.e., it rarely found a valid design for the wrong target). This suggests TB effectively prioritizes candidates with high posterior probability of satisfying the intended range.
• Case Study 1: Polymer Synthesis (KMC): In a task to reproduce specific molecular weight distributions (MWD) via reaction conditions, TB recovered valid designs significantly faster and more consistently than other methods. It identified diverse reaction pathways (degenerate solutions) that produced similar MWDs, offering flexibility for process optimization.
• Case Study 2: Oligomer Discovery: In a discrete library of sequence-defined conjugated oligomers, TB achieved the highest mean diversity score. It successfully identified structurally distinct molecules that produced nearly identical optical absorption bands, demonstrating the ability to navigate structure-property degeneracy.

5. Significance and Claims

The paper claims that range-aware BO provides a practical and sample-efficient foundation for specification-driven design, particularly when design flexibility and solution diversity are critical.

• Alignment with Design Needs: The authors argue that TB aligns better with real-world design constraints where multiple valid options are preferred over a single "optimal" point, and where candidates may differ in cost, robustness, or processability.
• Superiority over Extremum Seeking: The results suggest that improvement-based or uncertainty-driven methods are often misaligned with range discovery, as they tend to drift toward off-target ranges or over-exploit promising regions. TB's explicit scoring of range satisfaction yields more reliable and diverse outcomes.
• Modest Scope: The authors note that while TB is a robust default choice, it is not universally the top-ranked method for every individual target or dataset. The performance gap between model-guided search and random sampling narrows when valid regions are dense in discrete libraries, indicating that the value of the surrogate model is highest when valid regions are sparse or disconnected.
• Future Directions: The paper suggests future work could involve integrating explicit diversity incentives into the acquisition, relaxing the isotropic variance approximation via multi-output GPs, and developing adaptive tolerances. It envisions the integration of this framework with automated synthesis pipelines for closed-loop materials discovery.