AI FOR SCIENCE · TIMELINE
The 5 entries the timeline holds for 2021, newest first. Each was read at source. The full timeline, with filters →
What happened: A new commercial company led by Demis Hassabis is created to rebuild drug discovery around an AI-first approach.
Why it matters: AlphaFold's scientific success produces a dedicated company aimed at therapeutic translation.
What happened: Andrij Vasylenko, Matthew J. Rosseinsky and colleagues at the University of Liverpool trained a variational autoencoder on 2,021 four-element phases in the Inorganic Crystal Structure Database and ranked unexplored combinations by how likely they are to hold new compounds. They chose Li-Sn-S-Cl, ranked fifth, partly for its parent compounds' conductivity, aimed synthesis with structure prediction, and made Li3.3SnS3.3Cl0.7. Its bulk conductivity at room temperature is 3.2 x 10^-5 S/cm.
Why it matters: The model helped pick which chemistry to try, and that pick did hold a new compound. It did not find a better conductor: the paper puts it slightly below known Li4SnS4, says researchers made the final call, and notes that a model trained only on compounds that exist cannot rule other choices out. The same team made and measured it.
What happened: The first release makes roughly 350,000 predicted structures searchable, including the human proteome.
Why it matters: Model inference becomes a public research resource that can be reused without rerunning the model.
What happened: The architecture, code and methods behind the CASP14 result become available to the research community.
Why it matters: The breakthrough starts turning into shared scientific infrastructure rather than a single demonstration.
What happened: Christopher M. Collins, Matthew J. Rosseinsky and colleagues mapped Y-Sr-Ti-O and Ba-Y-Ti-O with computed probe structures and used random-forest and neural-network models of thermal conductivity to pick a region predicted near 7 W/mK, against about 16 elsewhere. There they made Ba10Y6Ti4O27, an aperiodic but ordered titanate, whose measured thermal conductivity at 298 K is 0.715 W/mK.
Why it matters: The models pointed at a region, not at this compound: the paper says its aperiodic structure is unlike anything in the training set, so the record low came from the chemists' search inside the region the models chose. The same team made and measured it, and low heat flow is a property heat barriers and thermoelectrics need.