Artificial intelligence (AI) is often used to study medical images, materials, and scientific data. However, many AI systems have trouble when the data comes from real-world sources that differ in clarity, errors, or trustworthiness. For example, information from various tools or tests can have different levels of detail or noise. Older AI models usually ignore these differences, leading to less accurate outcomes.
To fix this, researchers created a new AI system called ZENN, short for Zentropy-Embedded Neural Networks. This system trains AI to notice and adjust for hidden variations in data quality by combining rules from quantum mechanics, thermodynamics, and statistical mechanics.
The researchers built ZENN by adding these thermodynamic ideas directly into neural networks, a kind of AI that works like the human brain by connecting many processing units.
How ZENN works
ZENN splits data into two parts: energy, which captures the important patterns, and intrinsic entropy, which handles the noise and uncertainty. It also uses a temperature setting that can be adjusted to spot differences, like whether data is from clean computer simulations or messier real experiments. This helps the AI focus on real signals while dealing with varying quality.
In tests, ZENN performed as well as bigger AI models but was more reliable with uneven data. It also explains why things happen, not just what the results are. For instance, in studying an iron-platinum alloy, a metal mix that shrinks when heated, ZENN mapped out its energy changes to show the hidden reasons behind this odd behavior.
This approach could help in medicine, like analyzing mixed data on Alzheimer's disease to find disease stages or key changes. In materials science, it bridges perfect simulations and actual tests for designing things like medical implants. It might even aid quantum computing, where uncertainty is built-in. While scaling up for huge systems is still a challenge, ZENN shifts AI from just spotting patterns to uncovering how things really work, advancing science.
This research is published in PNAS.