The screening rule that finally asked whether a molecule can be made

Materials discovery is usually sold as a speed problem. Spread enough candidate structures in front of a fast enough model, the story goes, and the good one will surface by itself. What actually blocks interface molecules for perovskite cells is not a shortage of candidates; it is that the requirements contradict one another.

A self-assembled monolayer sits between the transparent electrode and the perovskite, tasked with pulling holes out cleanly. Its energy levels must align, its anchoring to the substrate must hold, the molecules must not stack into a mess, the synthesis has to be walkable, and the whole thing must survive the solution processing that follows. Excelling at any one of these proves nothing, yet a conventional trial can only optimise one or two at a time.

The evaluation scheme was the deeper problem. Computational screens tend to rank whatever is easiest to calculate, such as energy levels or hole reorganisation energy, and leave synthetic difficulty for a chemist to discover at the bench. The result is a leaderboard full of molecules that cannot be made, or cannot be made pure. Between accurate calculation and a real flask there is a stretch nobody owns.

This framework pulls that stretch into round one. The team extracted 308 functional fragments from carbazole derivatives in a public compound library, spanning more than 90,000 possible combinations, chose 24 fragments for modular assembly, built 300 candidate molecules, and used an energy-level filter to keep 124.

The decisive move came next. Synthesizability, molecular energy levels, hole reorganisation energy, aromaticity and heteroatom content, eight descriptors in all, were folded into a single ranking. A molecule was no longer judged only on its behaviour inside an ideal model but on whether it could be built reliably. That turns screening from a physics exercise into an engineering one.

Eight descriptors sounds like parameter piling, but the real work is assigning weights to goals that fight each other. Level alignment often trades off against ease of synthesis; a rigid aromatic core buys stability at the cost of solubility. Optimising a single number was never going to work. What the ranking returns is a set of acceptable compromises, and crucially, ones an experiment can check.

The validation is tidy. Three high-ranked molecules were synthesised, and the measured device performance followed the same order as the eight-descriptor scores. For a screening framework that matters more than any single record: it suggests the scoring carries transferable predictive power rather than having got lucky once.

In performance terms, the top molecule reached 27.2% power conversion efficiency. The more telling figure is durability: under 65 degrees Celsius, continuous illumination and maximum power point tracking, the device retained roughly 95% of its initial efficiency after 2,000 hours. Efficiency and lifetime standing together is what gets a result past the first gate between a paper and a production line.

That 2,000 hours needs to be read in the right frame. It is not equivalent to twenty-five years outdoors; accelerated ageing and real weathering still require a conversion that carries uncertainty. What it does is turn a long-standing worry about whether the interface layer fails first from speculation into a measured curve.

For me the methodological point outweighs the device. The model does not replace the chemist here. It compresses expert knowledge, high-throughput calculation and experimental constraints into one scoring rule. The intelligent part of intelligent screening is knowing which costs to price before anyone starts optimising.

Because the framework is modular, a different fragment library and a different descriptor set could in principle be dropped in for another class of interface materials. That is worth more than any single molecule it delivered. Beyond photovoltaics, any interface problem that must juggle performance with manufacturability runs into the same multi-objective tug of war.

The limits are real. A public library defines the ceiling of the search space, the weights still carry a degree of empirical judgement, and three verified molecules is not a large sample. A framework can shrink the area that has to be tried; it cannot guarantee the optimum lies inside the area it kept.

Adding synthesizability to a scorecard looks like one extra column. In practice it is a change of posture for computational materials work: from pursuing the optimum inside a model to admitting the constraints on a bench. The grand-sounding 90,000 combinations are the least interesting part. What matters is that three test tubes actually made it under the lamp.