Unified framework for adaptive modeling with density increasing feedback
A unified framework has been designed to create and maintain a set of adaptable general models that can be deployed and efficiently trained to fit to various deployments. The unified framework incrementally fills the feature space of a high dimensionality training dataset with field observations to reduce sparseness, trains and retrains a model set with the changing global training dataset, and then deploys a selected adaptable general model for training/fitting to a specified deployment scenario. Data that is generated by deployment adapted models can be validated and then added to the global training dataset that is used to train and update the general models. With the increasing density of the global training dataset, the general models can more quickly converge for a deployment scenario.
1 . A method comprising:
obtaining a first training dataset that populates a first subset of a high dimensionality feature space, wherein the first training dataset is based on measurement logs;
training a first model using the first training dataset to obtain a first trained model;
deploying the first trained model to a plurality of different deployment scenarios, wherein each deployment scenario is associated with a different target type;
training the first trained model using a second training dataset to obtain a second trained model, wherein the second training dataset comprises field observations associated with a first deployment scenario of the deployment scenarios, wherein the second training dataset populates a second subset of the high dimensionality feature space;
generating, by the second trained model, first predictions based on measurements associated with a formation, wherein the measurements are associated with a first target type;
determining that a first prediction of the first predictions is valid and a second prediction of the first predictions is invalid, wherein said determining that the first prediction of the first predictions is valid and the second prediction of the first predictions is invalid comprises determining that the first prediction satisfies a criterion that triggers updating of the first training dataset and the second prediction does not satisfy the criterion;
in response to determining that the criterion is satisfied by the first prediction, updating the first training dataset to include the first prediction but not the second prediction; and
training a third model using the updated first training dataset to obtain a third trained model.
2 . The method of claim 1 , wherein training the first model comprises training the first model for a target type corresponding to a petrophysical parameter.
3 . The method of claim 1 , wherein the measurement logs comprise synthetic measurement logs.
4 . The method of claim 3 , wherein the measurement logs further comprise historical field observations.
5 . The method of claim 1 , wherein the measurement logs comprise nuclear measurement logs.
6 . A non-transitory, machine-readable media comprising program code stored thereon, the program code comprising program code to:
obtain a first training dataset that populates a first subset of a high dimensionality feature space, wherein the first training dataset is based on measurement logs;
train a first model using the first training dataset to obtain a first trained model;
deploy the first trained model to a plurality of different deployment scenarios, wherein each deployment scenario is associated with a different target type;
train the first trained model using a second training dataset to obtain a second trained model, wherein the second training dataset comprises field observations associated with a first deployment scenario of the deployment scenarios, wherein the second training dataset populates a second subset of the high dimensionality feature space;
generate, by the second trained model, first predictions based on measurements associated with a formation, wherein the measurements are associated with a first target type;
determine that a first prediction of the first predictions is valid and a second prediction of the first predictions is invalid, wherein the program code to determine that the first prediction of the first predictions is valid and the second prediction of the first predictions is invalid comprises program code to determine that the first prediction satisfies a criterion that triggers updating of the first training dataset and the second prediction does not satisfy the criterion;
responsive to a determination that the criterion is satisfied by the first prediction, update the first training dataset to include the first prediction but not the second prediction; and
train a third model using the updated first training dataset to obtain a third trained model.
7 . The non-transitory, machine-readable media of claim 6 , wherein the program code to train the first model comprises program code to train the first model for a target type corresponding to a petrophysical parameter.
8 . The non-transitory, machine-readable media of claim 6 , wherein the measurement logs comprise synthetic measurement logs.
9 . The non-transitory, machine-readable media of claim 8 , wherein the measurement logs further comprise historical field observations.
10 . The non-transitory, machine-readable media of claim 6 , wherein the measurement logs comprise nuclear measurement logs.
11 . An apparatus comprising:
a processor; and
a machine-readable medium having program code executable by the processor to cause the apparatus to:
obtain a first training dataset that populates a first subset of a high dimensionality feature space, wherein the first training dataset is based on measurement logs;
train a first model using the first training dataset to obtain a first trained model;
deploy the first trained model to a plurality of different deployment scenarios, wherein each deployment scenario is associated with a different target type;
train the first trained model using a second training dataset to obtain a second trained model, wherein the second training dataset comprises field observations associated with a first deployment scenario of the deployment scenarios, wherein the second training dataset populates a second subset of the high dimensionality feature space;
generate, by the second trained model, first predictions based on measurements associated with a formation, wherein the measurements are associated with a first target type;
determine that a first prediction of the first predictions is valid and a second prediction of the first predictions is invalid, wherein the program code to determine that the first prediction of the first predictions is valid and the second prediction of the first predictions is invalid comprises program code to determine that the first prediction satisfies a criterion that triggers updating of the first training dataset and the second prediction does not satisfy the criterion;
responsive to a determination that the criterion is satisfied by the first prediction, update the first training dataset to include the first prediction but not the second prediction; and
train a third model using the updated first training dataset to obtain a third trained model.
12 . The apparatus of claim 11 , wherein the program code to train the first model comprises program code to train the first model for target type corresponding to a petrophysical parameter.
13 . The apparatus of claim 11 , wherein the measurement logs comprise synthetic measurement logs.
14 . The apparatus of claim 13 , wherein the measurement logs further comprise historical field observations.