Generative hierarchical searching for compounds

1 . A method performed by one or more computers, the method comprising:

receiving a natural language query specifying requirements for a compound;

processing the natural language query using a language policy to generate a plurality of representations of candidate compounds that each satisfy at least a subset of the requirements specified in the natural language query, each representation specifying at least a chemical formula of the corresponding candidate compound;

for each representation in a subset of the representations, using a generative machine learning model conditioned on the representation to generate one or more candidate chemical structures, each candidate chemical structure comprising a respective spatial location for each of the atoms of the corresponding candidate compound; and

selecting a chemical structure and corresponding compound from the plurality of candidate chemical structures.

2 . The method of claim 1 , wherein the subset of the representations is selected using a first heuristic function that determines a respective likelihood that each of the representations satisfies the requirements specified in the natural language query.

3 . The method of claim 2 , wherein the first heuristic function comprises a pre-trained language model.

4 . The method of claim 3 , wherein the pre-trained language model is prompted to rank the representations according to the respective likelihood that each of the representations satisfies the requirements specified in the natural language query.

5 . The method of claim 1 , wherein the selected chemical structure and corresponding compound is selected from the plurality of candidate chemical structures using a second heuristic function, the selection being based on a prediction of one or more properties of each candidate chemical structure.

6 . The method of claim 5 , wherein the second heuristic function comprises a machine learning model that has been trained to process an input comprising a chemical structure to generate an output comprising a prediction of one or more properties of the chemical structure.

7 . The method of claim 6 , wherein the one or more properties of the chemical structure comprise one or more of: a formation energy; a bandgap energy; a conductivity; a magnetic property; an electrical property; a mechanical property; and a phase change property.

8 . The method of claim 7 , wherein the machine learning model is a graph neural network.

9 . The method of claim 1 , wherein selecting the chemical structure and corresponding compound from the plurality of candidate chemical structures comprises performing a tree search over the representations and candidate chemical structures.

10 . The method of claim 9 , wherein the tree search uses one of: a best-first search strategy; a breadth-first search strategy; and a depth-first search strategy.

11 . The method of claim 1 , wherein the language policy comprises a language model that processes the natural language query to generate the plurality of representations of candidate compounds.

12 . The method of claim 11 , wherein the language policy comprises a retrieval policy that processes the natural language query to retrieve data relating to the natural language query from one or more repositories, and wherein the language model is conditioned on the data retrieved from the one or more repositories.

13 . The method of claim 12 , wherein the data retrieved from the one or more repositories is added to the natural language query before the language model processes the natural language query.

14 . The method of claim 1 , wherein each representation identifies a point group or space group for the corresponding candidate chemical structure.

15 . The method of claim 1 , wherein each representation is a text representation or a multimodal representation that comprises text data.

16 . The method of claim 1 , wherein the natural language query comprises a plurality of known chemical formulas and the requirements specify that the compound is not one of the plurality of known chemical formulas.

17 . The method of claim 1 , wherein the requirements specify a chemical family for the compound.

18 . The method of claim 1 , wherein the requirements specify a stoichiometry or range of stoichiometries for the compound.

19 . One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:

receiving a natural language query specifying requirements for a compound;

processing the natural language query using a language policy to generate a plurality of representations of candidate compounds that each satisfy at least a subset of the requirements specified in the natural language query, each representation specifying at least a chemical formula of the corresponding candidate compound;

for each representation in a subset of the representations, using a generative machine learning model conditioned on the representation to generate one or more candidate chemical structures, each candidate chemical structure comprising a respective spatial location for each of the atoms of the corresponding candidate compound; and

selecting a chemical structure and corresponding compound from the plurality of candidate chemical structures.

20 . A system comprising:

one or more computers; and

one or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform operations comprising:

receiving a natural language query specifying requirements for a compound;

processing the natural language query using a language policy to generate a plurality of representations of candidate compounds that each satisfy at least a subset of the requirements specified in the natural language query, each representation specifying at least a chemical formula of the corresponding candidate compound;

for each representation in a subset of the representations, using a generative machine learning model conditioned on the representation to generate one or more candidate chemical structures, each candidate chemical structure comprising a respective spatial location for each of the atoms of the corresponding candidate compound; and

selecting a chemical structure and corresponding compound from the plurality of candidate chemical structures.