The End of Serendipity: How Inverse Design in Materials Informatics Is Turning Materials Design on Its Head

In game design, for which I am a bit of a nerd, there is a well-known concept related to the interplay between two design methodologies: ‘top down’ and ‘bottom up’.

In short: ‘top down’ design starts with a thematic or conceptual layer, and works out game mechanics, features and content flowing from that.  On the other hand, ‘bottom up’ design begins with the details: mechanics, operational actions a player can take. Conceptual or thematic layers are then constructed based on and around that framework.

Both methodologies have their benefits and weaknesses; marrying the two is a tricky business. But when it works, it produces fantastic results.

I’ve always thought the same applicable in other fields. One standout is that of materials design.

From Discovery to Design the ‘bottom up’ approach

Picture the scene. A new alloy, polymer, or ceramic is synthesised. Researchers measure its properties and investigate its behaviour. If those properties prove useful, applications emerge. If not, the material may remain little more than an interesting academic result.

This approach reflects a fundamental reality: the number of possible materials is vast. Even when restricted to chemically plausible compounds, the space of potential compositions and structures is effectively limitless. When something new and interesting is found, we naturally want to make the most of it.

As a consequence, discovering useful materials has often resembled searching for a needle in a haystack. Progress depended on scientific intuition, experimental skill and, occasionally, serendipity.

At many points in the history of materials science, innovation has followed this ‘bottom up’ approach. A new material is discovered, its properties are characterised, and then researchers begin looking for commercial uses.

That is, we start at the ‘detail’ stage, and then work up in the direction of ‘application’.

Sometimes this process produces transformative technologies. Teflon was discovered accidentally while investigating refrigerants. Graphene was isolated through experiments that initially seemed almost playful. In each case, the material came first. Understanding what it could be used for came later.

This ‘material first’ approach has delivered extraordinary technological progress. Yet it can also shape the way we think about materials innovation. We often assume that discovery of a new material is the beginning of the journey, and the end is the product.

Changing Direction – inverse design and the ‘top down’ approach

Increasingly, materials informatics is letting researchers shift methodologies more heavily to a ‘top down’ approach.

In this context, a ‘top down’ approach involves first defining the problem that is to be solved, or the properties that are desired, and then finding materials which meet those requirements.

This inverse design strategy has garnered much interest over the last few years. Going back to 2018, Zunger described “a new style of collaboration between theory and experiment is discussed, whereby the desired functionality of the new material is declared first and theoretical calculations are then used to predict which stable and synthesizable compounds exhibit the required functionality”.

This defines the basis of the ‘top down’ approach - work out what you want, and then find materials that satisfy the functions needed.

This is of course an approach with a long history of its own: however, searches for materials are often constrained by existing paradigms. Often developments of this type begin with a known ‘best in class’ and look to improve it.

The Materials Genome Vision

A major milestone came in 2011 with the launch of the Materials Genome Initiative (MGI) in the United States. 

The ambitious objective was simple: halve the time and cost required to move new materials from discovery to commercial deployment. Rather than relying primarily on experimentation, the initiative sought to integrate computation, data infrastructure and experimental validation into a more predictive development process.

The initiative reflected a growing recognition that materials discovery had become a bottleneck in technological progress. Whether the challenge was batteries, semiconductors, catalysts or structural materials, the pace of materials innovation increasingly constrained advances elsewhere.

One of the most influential outcomes of this movement was the Materials Project, launched at Lawrence Berkeley National Laboratory. By combining high-throughput computational methods with openly accessible materials data, the project demonstrated how large-scale databases could support a fundamentally different approach to materials research.

Instead of asking researchers to search manually through decades of literature and experimentation, the idea was to create a searchable map of materials space.

For the first time, the foundations existed for materials discovery at digital scale.

The Logic of Inverse Design

The intellectual leap from traditional ‘discovery’ (bottom up) to inverse design (top down) is easy to describe. A bottom up methodology asks: given this material, what properties does it have? A top down one asks instead: given these desired properties, what material should exist?

Consider a material such as a hydrogen-separation membrane.

Historically, researchers might screen known materials, test candidate formulations and gradually improve performance through iterative experimentation.

Under an inverse design framework, the process begins with a specification of properties:

• Selectivity

• Operating temperature

• Chemical stability

• Mechanical strength

• Manufacturing cost

• Environmental constraints

• Etc.

These characteristics become the input. The material becomes the output.

Why Machine Learning Matters

The growing interest in artificial intelligence within materials science is often framed in terms of speed. Machine-learning systems can evaluate candidate materials more quickly than conventional computational methods, allowing researchers to examine larger numbers of possibilities.

Speed matters, but a potentially more important leg is the implementation of ‘inverse design’. Machine learning enables researchers to navigate regions of materials space that may simply be too large and complex for traditional approaches.

Wang et al argued that machine-learning methods offer the potential to identify structure-property relationships that may be difficult, or even impossible, to uncover using conventional methodologies alone.

This is particularly important because modern materials challenges rarely involve optimising a single characteristic.

Battery materials, for example, must balance energy density, durability, charging speed, safety, cost and supply-chain constraints. Catalysts must balance activity, selectivity and long-term stability. Polymers increasingly need to satisfy competing requirements around performance, recyclability and sustainability.

Exploring these trade-offs manually becomes increasingly difficult. Machine-learning systems provide a mechanism for identifying promising regions of the search space and focusing human attention where it is most valuable.

In other words, they do not merely accelerate discovery. They reshape how discovery is performed.

Designing Materials That Do Not Yet Exist

Perhaps the most striking recent example of this new paradigm comes from Google DeepMind's Graph Networks for Materials Exploration (GNoME).

Published in Nature in 2023, the work combined graph neural networks with active-learning approaches to explore vast areas of crystalline materials space.

The headline figures attracted considerable attention. The system identified millions of candidate crystal structures and hundreds of thousands predicted to be thermodynamically stable.

Whether all of these candidates ultimately prove useful is, in many ways, beside the point. What matters is what the work represented.

Historically, researchers could only explore tiny fractions of the available design space. Systems such as GNoME suggest a future in which computational models can survey enormous regions of chemical possibility, identifying candidates that satisfy predefined performance objectives before any physical experiment is conducted.

The question changes from: “which materials should we test?” to “which materials should exist?”. That is a fundamentally different way of thinking about discovery.

Beyond Prediction

Despite the excitement surrounding inverse design, important challenges remain.

Materials science differs from many other AI application areas because the underlying data are often sparse, inconsistent and fragmented. Experimental conditions vary widely. Negative results frequently remain unpublished. Manufacturing processes can influence performance as much as material composition itself.

Moreover, computational predictions are not products. Even the most promising candidate material must still be synthesised, characterised, manufactured and validated under real-world conditions.

This reality explains the growing interest in autonomous experimentation and so-called "self-driving laboratories". Rather than attempting to replace scientists, these systems seek to integrate prediction, synthesis, testing and learning into increasingly sophisticated feedback loops.

More than just a game

The most interesting aspect of materials informatics may not be that it allows us to move faster. It may be that it encourages us to think differently about what discovery actually means.

Understanding and implementing innovation in a ‘top down’ rather than a ‘bottom up’ is a fascinating shift. Marrying those two together, as we’ve seen in game design, may be the key to the biggest successes yet.