Image Credit: Scientific Frontline / stock image
Scientific Frontline: Extended "At a Glance" Summary: Physics-Informed AI in Drug Delivery
The Core Concept: Physics-informed neural networks (PINNs) are artificial intelligence models pre-programmed with fundamental physical laws to accurately predict how quickly controlled-release materials will dispense therapeutic agents.
Key Distinction/Mechanism: Unlike standard AI models that rely entirely on massive datasets to identify patterns, PINNs integrate short-term experimental observations with known physical principles. For simple planar materials, this reduces the required experimental data to just 6%, effectively cutting laboratory testing time by 94%.
Major Frameworks/Components:
- Physics-Informed Neural Networks (PINNs): The underlying AI architecture that embeds physical laws directly into the machine learning algorithm to drastically reduce training time and data dependency.
- Fick's Law of Diffusion: The primary physical principle utilized in this model, describing the migration of molecules from areas of high concentration to areas of lower concentration.
- Bayesian Statistics: An additional mathematical layer integrated into the neural network to quantify uncertainty and manage noisy laboratory data, ensuring highly precise predictive outputs.
Branch of Science: Artificial Intelligence, Computational Pharmaceutics, Applied Mathematics, and Materials Engineering.
Future Application: The rapid development and optimization of external therapeutic patches, smart bandages, internal biomedical implants, and oral controlled-release medication systems.
Why It Matters: By drastically reducing the need for iterative, time-consuming physical experiments, this approach accelerates the pharmaceutical development pipeline, bringing advanced drug delivery systems to market faster and more cost-effectively.
Brown University researchers have developed a new artificial intelligence method for predicting the rate at which materials used in controlled drug-release systems will release therapeutic agents.
The new method could slash the development time for new therapeutic patches, bandages, and implants.
“The current methodology for developing controlled-release materials is based on experiment,” said Vikas Srivastava, an associate professor of engineering at Brown. “You design a material, test it with experiment, tweak the design, and experiment again. That takes a lot of time. What we’ve developed is a way of using physics-informed neural networks to make accurate predictions with only a little bit of data, which can save a huge amount of time in developing new drug delivery systems.”
In a study published in the Journal of Drug Delivery Science and Technology, Srivastava and his colleagues tested an approach that uses physics-informed neural networks (PINNs) to make predictions about the properties of candidate materials. The standard neural networks used in AI models require extensive training and mounds of data to produce accurate predictions. But PINNs—which were originally developed by Brown mathematician and professor George Karniadakis—start with fundamental physical laws baked into the system. That reduces training time and enables the models to return accurate prediction results with little or no training data.
For the study, Srivastava worked with Daanish Qureshi, who received his bachelor’s degree from Brown in 2025, and Khemraj Shukla, an associate professor (research) of applied mathematics. The team developed PINNs equipped to combine short-term experimental observations with Fick’s law of diffusion, which describes how molecules migrate from areas of high concentration to lower concentration, to predict the long-term behavior in the controlled release of drugs.
Using existing experimental data on various controlled-release materials, the researchers tested how much data was needed for the PINNs to make predictions that matched the real-world experimental outcomes. They found that the PINNs required only the first 6% of the experimental data to return accurate predictions for simple, planar materials. For more complex materials—those with folds or wrinkles—the PINNs required 33% of the experimental data.
“We’re basically cutting the time required for experiment by 94% for simple materials and 67% for more complex ones,” said Srivastava, who works with the Institute for Biology, Engineering, and Medicine at Brown. “In pharmaceutical development, time is money. We’re hopeful that this approach can help in getting products to patients more quickly and less expensively.”
To improve the output of the models further, the researchers supplemented the diffusion model PINN with a version that includes Bayesian statistics. Even controlled laboratory experiments involve a bit of uncertainty and noisy data. The Bayesian PINNs are able to quantify that uncertainty to produce output that more precisely reproduces experimental data.
While the research in this case focused on the materials used for external patches and bandages, the same general basic concepts apply to pills and other forms of controlled-release systems. Srivastava says the basic approach demonstrated here could be useful for those systems as well.
“We believe this demonstrates an area in which AI can make a real difference in developing products that improve people’s lives,” Srivastava said.
Published in journal: Journal of Drug Delivery Science and Technology
Title: Drug release modeling using Physics-Informed Neural Networks
Authors: Daanish Aleem Qureshi, Khemraj Shukla, and Vikas Srivastava
Source/Credit: Brown University
Edited by: Scientific Frontline
Reference Number: ai070626_01