Artificial intelligence, they call it — not that the intelligence is somehow fake. It’s true intelligence, but humans also make it. That implies that AI — a power tool that can add speed, effectiveness, insight and precision to the work of a researcher — has many limitations.
It is only as strong as the techniques and knowledge it has been provided. It does not know, on its own, what information is lacking, how much weight it takes to provide various kinds of information, or whether the information it relies on is inaccurate or corrupt. It does not deal with confusion or random events precisely — unless it knows how. It does not use the expertise experts have acquired over years and physical models that underpin physical and chemical phenomena, relying solely on data, as machine-learning models traditionally do. Teaching the computer to organize and incorporate data from widely differing sources has been challenging.
Researchers at the University of Delaware and the University of Massachusetts-Amherst have now published descriptions of a new approach to artificial intelligence, which in its calculations builds complexity, error, physical rules, expert information and missing data and eventually leads to far more accurate models. The new approach offers assurances that usually lack AI models, illustrating how useful the model can be for achieving the desired outcome.
Joshua Lansford, a PhD student in the Department of Chemical and Biomolecular Engineering at UD, and Prof. Dion Vlachos, director of the Energy Innovation Catalysis Center at UD, are co-authors of a paper published in the journal Science Advances on Oct. 14. Jinchao Feng and Markos Katsoulakis of the University of Massachusetts-Amherst Department of Mathematics and Statistics also contributed.
For computer models used in many fields of science, the new mathematical paradigm may generate higher performance, accuracy and innovation. These models provide powerful ways, instead of in the laboratory, to analyze data , research materials and complex interactions and tweak variables in virtual ways.
“Traditionally, in physical modeling, we first construct a model using only our physical intuition and device skills,” Lansford said. “Then after that, because of errors in underlying variables, we calculate uncertainty in forecasts, mostly depending on brute-force methods, where we sample, then run the model to see what happens.”
Successful, accurate models save time and money and point to more effective techniques , new materials, higher accuracy and novel methods that might not otherwise be considered by researchers.
In a chemical reaction known as the oxygen reduction reaction, the paper explains how the latest mathematical system operates, but it is applicable to several forms of modeling, Lansford said.
He said, “The chemistries and materials that we need to make things quicker or even make them feasible, such as fuel cells, are extremely complex.” We need consistency …. And you need to have bounds on your prediction error if you want to make a more successful catalyst. You will tighten the field to explore by intelligently determining where to put your efforts.
“In the nature of our model, uncertainty is accounted for,” Lansford said. “Now it is no longer a model of determinism. It’s a probabilistic one.’
The model itself defines what knowledge is required to minimize model errors with these latest mathematical advances in place, he said. Then it is possible to use a higher level of theory to create more detailed results or to generate more results, leading to even smaller error limits on the predictions and reducing the field to be explored.
“These estimates take time to produce, so we always work with tiny datasets—10-15 data points. That’s where the apportionment error needs to come in.”
That’s still not a money-back guarantee that the commodity needed will be produced precisely by using a particular material or approach. But it’s a far closer guarantee than you would have gotten before.
This new model design approach could dramatically boost work in renewable energy, battery technology, mitigation of climate change, drug discovery, astronomy, economics, physics , chemistry and biology, just to name a few examples.
Artificial intelligence doesn’t mean that it no longer requires human experience. The absolute contrary.
For any computational model, the expert expertise that emerges from the laboratory and the rigors of scientific investigation is central, fundamental content.
Source of Story: University of Delaware Materials Supplied. Published in the original by Beth Miller. Note: For style and length, material can be edited.
No CommentsLeave a comment Cancel