Although artificial intelligence (AI) can adapt to changing conditions and achieve desired outcomes, how algorithms “understand” and adjust to inputs can be a mystery.

Lin et al. sought to uncover this “black box” in AI-controlled cold atmospheric plasma (CAP) treatments, an approach that induces apoptosis in diseased cells while preserving healthy ones. In previous work, they developed a machine learning (ML) system that predicts the post-treatment state of cancer cell targets and adjusts treatment accordingly. However, they didn’t know how the ML system achieved this outcome without an understanding of specific plasma parameters.

Using an AI-based optical emission spectroscopy (OES) spectra translation algorithm, the authors reverse engineered real-time chemical accumulations above cell medium surfaces. They found that, despite changing conditions, the ML algorithm alters experimental parameters to achieve the same therapeutic outcomes. The application of a Fourier transformation on OES spectra and chemical kinetics analysis revealed how the ML algorithm independently captured additional layers of physics information relying solely on cell viability status, without human input of this information, to achieve the precision and reliability of their AI-controlled CAP model.

“Beyond plasma medicine, similar approaches could advance machine learning-based control in fields like electric propulsion for satellites, plasma-based microfabrication, fusion reactor management, and many other plasma applications” said author Michael Keidar.

Next, the team looks to extend the scope of control that was demonstrated in this paper.

“Instead of limiting the AI to adjusting treatment duration, we plan to authorize and train the AI to control multiple plasma parameters simultaneously, including voltage, gas flow rate, and even additional external electric fields,” said author Li Lin. “In doing so, we aim to tailor therapy to the specific needs of each patient.”

Source: “Low-temperature plasma adaptation in the course of machine learning controls of plasma medicine,” by Li Lin, Qihui Wang, Zichao Hou, Michael Keidar, Physics of Plasmas (2025). The article can be accessed at https://doi.org/10.1063/5.0274614 .