Project SLAMD - Data Driven Discovery of Sustainable Building Materials

20/07/2022

• Challenge
• Methods
• Objectives
• Partners

Concrete represents the greatest mass flow of mankind - the production of its binder Ordinary Portland cement, however, leads to a very high CO2 emission. The development of climate-friendly alternatives is highly complex as they can have a wide range of composition strongly effecting its properties. What makes their development so difficult is that there are significantly more alternative binder compositions than normal cements - but only a few of them would stand up to the quality requirements in practice.

Sequential Learning ranks the experiments based on their utility and is frequently recognized as having great potential to accelerate materials research with a small number of highly complex data points. This is done by coupling the predictions of an ML model with a decision rule that guides the experimental procedure. The new app will provide flexible and low-threshold access to AI and enable the creation of various optimization scenarios for benchmarking AI performance.

With our work we develop and optimize the so-called Sequential Learning App for Materials Discovery ("SLAMD"). The idea is to use Artificial intelligence (AI) and Machine Learning (ML) to give building materials research a new dynamic. Particularly, the use of Sequential Learning (SL) promises a breakthrough by identifying promising experiments quickly and early on discard unfavorable experimental paths. This allows to develop new materials with significantly fewer experiments overall.

Project coordination

Bundesanstalt für Materialforschung und -prüfung (BAM)
Department Non-Destructive Testing

Partners

TU Berlin - Institute of Civil Engineering, Prof. Stephan
Iteratec - Software engineering

SLAMD - Sequential Learning App for Materials Discovery

We are investigating how to accelerate experimental materials science through a data-driven approach in the field of alternative, climate friendly building materials. Common Machine Learning (ML) models are used to predict final product properties based on compositional and processing data. Specifically, the task of ML is often to find materials with desired properties in a high-dimensional search in the discovery space (DS) spanned by many possible material mixtures, of which only a small fraction has been empirically validated.

The challenge for the ML model is to make the best use of the limited knowledge from the few available data points to effectively explore the DS. For a model to achieve good generalizability hundreds to thousands of training data are typically required. This makes it impractical for a discovery task where materials are new and not much data is yet available.

Sequential learning (SL) on the other hand, looks beyond the already empirically known materials for new end products and generally gets by with much less data. Instead of relying solely on predictions, the approach creates a feedback loop: the AI suggests potentially useful experiments which are then empirically validated in the lab. The validation data is used in the next round to make even better suggestions. This reduces the number of unsuccessful experiments, i.e., experiments that lead to materials with undesirable properties, such that an ideal (as small as possible) number of successive experiments is achieved.

The performance of SL can be demonstrated by simulated experiments. Here, the actual results are already known for all data points. However, only a small part of the data is provided to the SL algorithm (although more training data would be available). These data are extended in each iteration by a new data point from the rest of the available data. The SL algorithm decides for itself which next data point is most promising. The less data needed to find materials with the target properties, the more successful the optimization. This is tested repeatedly under random initial conditions. Success can then be measured in terms of the average number of required SL cycles.

Project coordination

Bundesanstalt für Materialforschung und -prüfung (BAM)
Department Non-destructive Testing

Project partners

TU Berlin - Institute for Structural Engineering, Prof. Stephan
Iteratec - Software engineering

Funding

The project is funded through internal BAM funds.