Our Partner Bundesanstalt für Materialforschung und -prüfung (BAM) announces the release of a new pre-print, “LLMs can Design Sustainable Concrete – a Systematic Benchmark.” Authored by researchers Christoph Völker, Dr. Tehseen Rug, Kevin Maik Jablonka, and Sabine Kruschwitz, this publication underscores the role of Large Language Models (LLMs) in material design, outshining traditional Data-Driven Design (DDD) methods in constructing materials.

This study is relevant to the building material design field and particularly impactful for the Reincarnate project’s digital strides towards circularity in the construction industry.

Paving the way for a more sustainable concrete solution

The study introduces a novel approach using ‘fuzzy design knowledge’, offering unmatched adaptability to varying conditions over conventional models. This methodology leads to the precise determination of optimal concrete compositions, exceptionally beneficial for circular materials. As 50% of what we produce is concrete, transitioning to sustainable practices is not just important, it’s essential. It adeptly addresses the challenges posed by the variability of waste streams, surpassing data inconsistency issues through the application of core design principles.

Unveiling the power of AI in real-world problem solving

A significant advancement in this research is the use of mathematical problem-solving strategies to refine AI-driven proposals. This approach not only improves the design process but also showcases AI’s capacity to resolve complex, real-world problems, extending beyond structured environments like mathematics.

Dr Christoph Völker sheds light on the process, “Our approach essentially orchestrated a discussion among AI bots, each informed with general design rules like ‘too much water reduces strength’ or ‘Fly ashes are less reactive than slags from steel production’. After generating concrete design proposals, these bots were then provided with real-world performance feedback (see image with roles). This process was automated and repeated under varying conditions for a reliable benchmark against our established design framework SLAMD (innovation 10). To our surprise, the chatbots not only generated viable concrete designs but also outperformed our baseline methods in reliability. This was unexpected, especially considering that LLMs do not use traditional training data, and it challenged our assumption that training data-based methods would be inherently more reliable.”

Dr Völker explains further, While we noticed initial improvements, there was a point where progress plateaued. To address this, we took inspiration from OpenAI’s work on enhancing chatbot problem-solving capabilities for text math problems. By evolving the chatbots’ interaction from simple question-answer exchanges (see image with simple loop) to more dynamic, iterative discussions (see image with extended loop), we achieved a significant breakthrough.  This advancement not only refined our design process but also demonstrated the extensive capability of AI-systems in complex problem-solving, far beyond the structured confines of purely mathematical environments.”

In summary, BAM’s research redefines the approach to material design in the construction industry and underscores the vast potential of AI in forging sustainable, adaptable solutions for the future.

 

 

Read the full paper here!

How about transforming our ‘concrete jungle’ into a more environmentally conscious and resilient material landscape? Inspired by these rhythms of change, we will continue to work on methods of material designs, construction products, components and buildings to harmonise with nature and promote circularity in the construction industry.

🔎 Key findings of this research

  • LLMs can design materials with no training data required – enabling the use of variable feedstocks in circular economy
  • LLMs outperform strong baseline methods – they are more reliable despite using fuzzy knowledge
  • LLMs capabilities can be scaled through widely adaptle methods – enabling them to solve more complex tasks.

 

Stay tuned to our research developments by subscribing to our newsletter and follow us on LinkedIn and Twitter so you don’t miss a thing!

 


  Read our previous articles on research works regarding materials: