AI in the Lab: The Future of BMD Testing Is Already Here

This guest post was written by Andrew B. Cooper.

There is no doubt that AI is becoming a part of our everyday lives. A quick look back at how things were just a few decades ago reveals the remarkable changes that have steadily transformed our daily routines. Self-driving cars, facial detection at airports, and the analysis of X-ray images by computer are all situations in which the ability of machines to detect and interpret features in images is now often more reliable than human judgment.

Asphalt mixtures today are far more complex than they were thirty or even twenty years ago. At the time PG grading was developed, asphalt mixtures consisted only of stone and neat bitumen. Today, they can include an array of different components: RAP, plastics, rubber, fibers, and more. Figuring out how all these combinations will perform tomorrow, or in a few years, using basic BMD (Balanced Mix Design) tests is asking a lot. As a result, although hugely popular, these tests can struggle to reliably predict the performance of some materials and mix types.

There is no collective desire to return to long, complicated, expensive tests such as those performed on the AMPT, but there is a clear need for tests that more accurately identify the behavior of multi-component mixtures.

Humans have five senses. When we encounter something new, we usually start by looking at it. Depending on what our brain tells us about those images, we might touch it, smell it, taste it, or listen to it. Our understanding is based on combining the relevant sources of feedback with prior knowledge and experience.

BMD (Balanced Mix Design) tests, however, base their results on just one source of feedback, the shape of the load curve. This isn’t because asphalt mixtures are best analyzed with only one "sense", but because these tests were built around the standard loading machines already found in most labs. In some labs, seasoned technicians can intuitively read crack patterns and predict performance. Unfortunately, that kind of insight is neither scalable nor easily transferable.

A more quantitative and consistent approach, free from individual bias, is possible with the use of digital image analysis alongside load curve data. BMD tests have been successfully analyzed in universities using DIC (Digital Image Correlation), but those setups are complex, expensive, and generally restricted to research settings.

MiAS (Materials image Analysis System) is a relatively new development that uses cameras and AI specifically for routine BMD tests. A key and significant advantage it holds over traditional DIC is its dedicated frame, which allows both sides of a specimen to be filmed during testing. With the MiAS Optical Flow for BMD software, it is possible to track the movement of points on the surface of both specimen sides. These displacements, combined with load curve analysis, produce more accurate, complete test outputs and help reduce COV. The system also produces a permanent video record of every test.

The MiAS captures localized deformations, crack initiation, crack paths, changes in horizontal diameter, the speed of deformation, and all of these measurements on both sides of each specimen.

Image analysis uses AI, and AI is improving at an astonishing rate. The multimodal capabilities of Large Language Models (LLMs), their ability to work across text, images, and now video, opens up huge potential to push these tests to a new level.

Imagine receiving feedback like:
“If  you reduce the RAP content by 10%, fracture energy is likely to increase by 12–18%, based on visual texture and historical data.”

Or entering an instruction such as:
“Design a mix for the IDEAL CT test that maximizes Gf but uses more than 20% RAP and PG 58-28 binder. Prioritize crack branching resistance.”

This level of interaction is coming very soon. Improve your mixes, reduce your COV, and be ready for AI in the lab.
MiAS is compatible with the Humboldt HM-5125A.3F, HM-5170A.3F, and  HM-5035.3F.