target property window is narrow. Low-loss PCB FRs must satisfy many constraints simultaneously: low Dk contribution, ultra-low Df contribution, FR efficiency, thermal stability, compatibility with CCL resin systems, processability, realistic synthesis routes, cost position and regulatory acceptability.
AI modelling is valuable because it can explore chemical space more broadly than a laboratory programme can synthesise and test. However, speciality chemical development often suffers from small datasets, inconsistent historical data and complex structure-property relationships. Unlike fields with very large training datasets, a niche target, such as ultra-low-loss FRs for advanced CCLs, may have only limited experimental data available.
The ICL – NobleAI collaboration addressed this by combining basic chemical understanding, computational chemistry, scientific AI and domain expertise. NobleAI’ s approach is described publicly as SBAI: models that incorporate scientific principles and domain knowledge rather than relying solely on statistical pattern recognition.
This is particularly relevant when datasets are limited, because chemistry-informed constraints can guide the model toward plausible structures and reduce the risk of generating molecules that are mathematically interesting but chemically unrealistic.
Focused discovery
The project began by defining the desired molecular profile. The target was not merely‘ a FR’, but one suitable for ultra-low-loss CCL and PCB applications. That required a combination of dielectric, thermal, chemical and practical properties.
Computational chemistry and chemical descriptors were used to represent relevant molecular features, while AI models explored candidate structures that could plausibly meet the target property set. NobleAI also provided dataenrichment services through patent data extraction that made it possible to train the SBAI models to calculate the set of target properties used to guide molecular generation.
ICL was able to generate more than 4,000 candidate molecules using NobleAI’ s platform. This stage allowed the team to move beyond the narrow set of structures traditionally considered in FR chemistry. At the same time, the initial output was only a starting point. A generated molecule is not automatically useful: it may be difficult to synthesise, unstable, commercially impractical, poorly compatible with the resin system or unsuitable for FR mechanisms.
ICL chemists reviewed the generated structures, assessed their plausibility, and helped refine the model. Molecules were filtered based on synthetic accessibility, expected FR mechanism, compatibility with low-loss resin systems, impurity risks, thermal behaviour and practical commercial considerations.
The custom model NobleAI developed for ICL was did not replace chemists; it was extended their reach and learning from their decisions. Through this iterative process, the original population of more than 4,000 molecules was narrowed to a much smaller and more relevant set. The model became more focused, and the chemical search space
70 SPECIALITY CHEMICALS MAGAZINE ESTABLISHED 1981