US 12,079,697 B2
Systems and methods for diversity-enhanced adaptive experimentation in a machine learning-based formulation platform
Michael L. Thompson, Cincinnati, OH (US); Manmit Shrimali, Oakville (CA); and Ajith Govind, Bengaluru (IN)
Assigned to Turing Labs, Inc., New York, NY (US)
Filed by Turing Labs, Inc., New York, NY (US)
Filed on Feb. 9, 2023, as Appl. No. 18/107,577.
Claims priority of provisional application 63/310,172, filed on Feb. 15, 2022.
Claims priority of provisional application 63/399,395, filed on Aug. 19, 2022.
Claims priority of provisional application 63/444,266, filed on Feb. 9, 2023.
Prior Publication US 2023/0259822 A1, Aug. 17, 2023
Int. Cl. G06N 20/00 (2019.01); G06N 5/022 (2023.01)
CPC G06N 20/00 (2019.01) [G06N 5/022 (2013.01)] 19 Claims
OG exemplary drawing
 
1. A computer-implemented method for adaptively generating experimental product formulations for accelerating a formulation design of a target physical product, the computer-implemented method comprising:
at a remote formulation service that is implemented by a network of distributed computers:
constructing, by one or more computer processors, a formulation objective function based on a plurality of distinct product-informative formulation parameters of the target physical product, wherein the plurality of distinct product-informative formulation parameters include:
(a) one or more product formulation objectives set for the target physical product;
(b) one or more formulation design variables associated with the target physical product; and
(c) one or more product formulation constraints imposed on the one or more formulation design variables associated with the target physical product;
iteratively generating, by the one or more computer processors via a probabilistic graphical model, distinct sets of product formulation parameter values based on the plurality of distinct product-informative formulation parameters of the formulation objective function;
automatically executing, by the one or more computer processors using a formulation simulation module, one or more computer-based formulation simulations that simulate a formulation performance of at least one distinct set of product formulation parameter values of the distinct sets of product formulation parameter values based on the probabilistic graphical model generating the at least one distinct set of product formulation parameter values, wherein each of the one or more computer-based formulation simulations enables a bypass of real-world experimentation of the at least one distinct set of product formulation parameter values;
automatically constructing, by the one or more computer processors a corpus of synthetic experimental findings data based on simulation output data of the one or more computer-based formulation simulations;
automatically adapting, by the one or more computer processors, the probabilistic graphical model using the corpus of synthetic experimental findings data that causes an intermediate model update of the probabilistic graphical model;
generating, by the one or more computer processors via the adapted probabilistic graphical model, an adapted set of product formulation parameter values that likely satisfy the one or more product formulation objectives thereby enabling an accelerated creation of the target physical product; and
generating, by the one or more computer processors, a target formulation proposal for creating the target physical product using the adapted set of formulation parameter values that likely satisfy the one or more product formulation objectives.