Method and system for augmented inversion and uncertainty quantification for characterizing geophysical bodies
A computer-implemented method for augmented inversion and uncertainty quantification for characterizing geophysical bodies is disclosed. The method includes machine-learning-augmented inversion that also facilitates the characterization of uncertainties in geophysical bodies. The method may further estimate wavelets without a well-log calibration, thereby enabling a pre-discovery exploration phase when well log data is unavailable. The machine learning component incorporates a priori knowledge about the subsurface and physics, such as distributions of expected rock types and rock properties, geological structures, and wavelets, through learning from examples. The methodology also allows for conditioning the characterization with the information extracted a priori about the geobodies, such as probabilities of rock types, using other analysis tools. Thus, the conditioning strategy may make the inversion more robust even when a priori distributions are not well balanced. Using the method, a scenario testing workflow may evaluate different candidate subsurface models, facilitating the management of uncertainty in decision-making processes.
1 . A computer-implemented method of performing geophysical inversion comprising:
performing a training step independent of measured data for a subsurface region to generate a machine-learned model, the training step using direct probabilistic modeling of geologic priors of at least one aspect of rock types, facies, or fluid types thereby tailoring the machine-learned model to the at least one aspect of rock types, facies, or fluid types;
accessing the measured data for the subsurface region;
solving an inversion problem by:
(i) inputting the measured data and an initial solution using the machine-learned model tailored in the at least one aspect to rock types, facies, or fluid types;
(ii) generating an inversion solution based on the measured data and the initial solution tailored in the at least one aspect to the rock types, the facies or the fluid types;
(iii) determining whether to continue iterating;
(iv) responsive to determining to continue iterating, iterating back to (ii) to generate the inversion solution that is updated and based on the measured data and the initial solution; and
responsive to determining not to continue iterating, drilling a well or extracting hydrocarbons from the well using the inversion solution tailored in the at least one aspect to the rock types, the facies or the fluid types.
2 . The method of claim 1 , wherein solving the inversion problem tailored in the at least one aspect to the rock types, the facies, or the fluid types comprises conditioning inversion based on scenarios of the rock types, the facies, or the fluid types; and
wherein the conditioning of the inversion is based on the scenarios of at least one of the rock types, the facies, or the fluid types obtained from seismic processing tools independent from priors based on knowledge bases and upstream to the inversion.
3 . The method of claim 1 , wherein solving the inversion problem tailored in the at least one aspect to at least one of the rock types, the facies or the fluid types comprises generating a function (f) that maps from latent space Z into parameter space of subsurface properties tailored to the rock types, the facies, or the fluid types.
4 . The method of claim 1 , further comprising accessing prior subsurface data;
wherein the prior subsurface data comprises rock types scenarios or facies scenarios;
wherein the machine-learned model comprises an augmented forward model based on machine-learning with the rock types scenarios or the facies scenarios; and
wherein solving the inversion problem is augmented based on the machine-learned model for the rock types or the facies scenarios.
5 . The method of claim 1 , wherein determining whether to continue iterating is based on the machine-learned model.
6 . The method of claim 5 , wherein the machine-learned model by:
comparing a solution generated by the machine-learned model with the inversion solution; and
determining whether to continue iterating based on the comparison.
7 . The method of claim 1 , wherein the measured data and the initial solution are separately input to generate the inversion solution.
8 . The method of claim 1 , wherein the machine-learned model generates a plurality of distinct and separate initial solutions;
wherein a plurality of iterative solutions are generated by solving the inversion problem based on the plurality of distinct and separate initial solutions; and
wherein the machine-learned model is used to evaluate the plurality of iterative solutions to determine whether to continue iterating.
9 . The method of claim 1 , wherein performing the training step consists of using training examples.
10 . The method of claim 9 , wherein the training examples comprise one or more rock type distributions or one or more geophysical property distributions.
11 . A computer-implemented method of machine learning-augmented geophysical inversion comprising:
accessing measured data for a subsurface region;
accessing prior subsurface data;
accessing conditioning data;
forming an augmented forward model independent of the measured data for the subsurface region based on:
a machine-learning model representing the prior subsurface data conditioned with the conditioning data, and
a physics model mapping output of the machine-learning model to seismic data thereby interlinking a direct physical relationship of at least one of rock types, facies, or fluid types with the seismic data;
initializing, using the augmented forward model, a plausible model by sampling latent space of a priori network model;
solving an inversion problem by:
(i) inputting the measured data and the plausible model using the augmented forward model tailored in the at least one aspect to rock types, facies, or fluid types;
(ii) generating, based on the measured data and the plausible model, one or more solutions which are consistent with the seismic data, the prior subsurface data and the conditioning data;
(iii) determining whether to continue iterating;
(iv) responsive to determining to continue iterating, iterating back to (ii) to generate the one or more solutions that is updated and based on the measured data and the plausible model; and
responsive to determining not to continue iterating, drilling a well or using the well to extract hydrocarbons using the one or more solutions.
12 . The method of claim 11 , wherein a distribution of latent space vectors is based on a normal distribution during the inversion, such that solving the inversion problem traverses within a subspace of high-resolution plausible subsurface models with augmentation from the α priori network model.
13 . The method of claim 11 , wherein solving the inversion problem based on multiple scenarios generates multiple solutions; and
wherein the machine-learning model evaluates the multiple solutions in order to rank the multiple scenarios.
14 . The method of claim 11 , wherein an augmented inversion method is combined with a nullspace search algorithm to find equally-likely or almost equally-likely solutions connected to the one or more solutions.
15 . The method of claim 11 , wherein the physics model associates geometric shapes of geobodies with geophysical or petrophysical properties; and
wherein the physical model associates shapes of channel systems and of lobe systems with their respective properties.
16 . A computer-implemented method of machine learning-augmented geophysical inversion comprising:
obtaining measured data for a subsurface region;
obtaining prior subsurface data;
performing a training step independent of the measured data for the subsurface region using direct probabilistic modeling of geologic priors of at least one aspect of rock types, facies, or fluid types to generate an augmented forward model based on machine-learning with the prior subsurface data conditional on scenarios of at least one of geologic systems, rock types, facies, or fluid types;
solving an inversion problem by:
(i) inputting the measured data and different rock distributions as different initial solutions using the augmented forward model, wherein different conditions of at least one of rock types, facies, or fluid types are used as inputs to the augmented forward model in order for the augmented forward model to generate the different rock distributions;
(ii) generating multiple scenario solutions based on the measured data and the different initial solutions;
(iii) determining whether to continue iterating;
(iv) responsive to determining to continue iterating, iterating back to (ii) to generate the multiple scenario solutions that are updated and based on the measured data and the different initial solutions;
responsive to determining not to continue iterating, testing at least one of the multiple scenario solutions; and
drilling a well or using the well to extract hydrocarbons using the at least one of the multiple scenario solutions.
17 . The method of claim 16 , wherein testing the at least one of the multiple scenario solutions comprises one of:
using the augmented forward model in order to evaluate the multiple scenario solutions; or
evaluating a loss function using the measured data for a subsurface region.
18 . The method of claim 16 , wherein the at least one of the multiple scenario solutions comprises a model; and
further comprising:
creating a map of categorical labels or probabilities of scenarios for each pixel or element in the model; and
creating, by the augmented forward model, examples that match the labels.
19 . The method of claim 16 , wherein uncertainty of the multiple scenario solutions at each spatial location is evaluated by optimization.
20 . The method of claim 16 , wherein uncertainty of the at least one of the multiple scenario solutions is evaluated by at least one of:
latent space projection and interpolation; or
extreme bounds analysis.
21 . A computer-implemented method of machine learning-augmented geophysical inversion comprising:
obtaining measured data for a subsurface region;
obtaining prior subsurface data;
forming an augmented forward model based on machine-learning with the prior subsurface data conditional on scenarios of at least one of geologic systems, rock types or facies and on a physics model mapping output of the augmented forward model to seismic data thereby interlinking a direct physical relationship of at least one of rock types, facies, or fluid types with seismic data;
using an augmented inversion method combined with a Hamiltonian null-space exploration problem to find multiple equally plausible solutions, wherein the augmented inversion method comprises:
(i) inputting the measured data and different rock distributions as different initial solutions
(ii) generating multiple scenario solutions based on the measured data and the different initial solutions, wherein the Hamiltonian null-space exploration problem using the multiple scenario solutions in order to find the multiple equally plausible solutions;
drilling a well or using the well to extract hydrocarbons using the at least one of the multiple equally plausible solutions.
22 . A computer-implemented method for simultaneously estimating wavelets and geophysical or petrophysical properties, the method comprising:
performing a training step independent of measured data for a subsurface region to generate an augmented forward model based on machine-learning with wavelet examples in order to determine representations in latent space;
accessing the measured data for a subsurface region;
solving an inversion problem by:
(i) estimating, based on different conditions input to the augmented forward model and without a well-log calibration, a plurality of initial wavelets;
(ii) estimating, based on the measured data, initial geophysical or petrophysical properties;
(iii) generating an inversion solution by simultaneously updating the plurality of initial wavelets and the initial geophysical or petrophysical properties;
(iv) determining whether to continue iterating;
V responsive to determining to continue iterating, iterating back to (iii); and
drilling a well or using the well to extract hydrocarbons using the inversion solution.
23 . The method of claim 22 , wherein the initial wavelets are estimated based on statistical wavelet estimations, elastic FWI results, or transferred from overlapping or nearby seismic surveys with well ties; and
wherein the initial wavelets comprise a distribution of wavelets as a prior to infer wavelets during the solving of the inversion problem.
24 . The method of claim 22 , wherein solving the inversion problem comprises inverting for wavelets by updates in the latent space of a wavelet network.