Magnetic resonance multi-scale focusing inversion method and system suitable for salt lake brine detection

Abstract

The application belongs to the field of geophysical exploration of salt lake brine resources, and is a magnetic resonance multi-scale focusing inversion method and system suitable for salt lake brine exploration, constructs an inversion model vector and a predicted magnetic resonance signal; constructs an inversion objective function formed by superimposing an observation response functional and a joint regularization term; converts the inversion objective function into an approximately derivable form, iteratively solves by using a momentum gradient descent method until the relative change of model data fitting difference meets a termination condition, and outputs a final inversion model vector; extracts a water content and a transverse relaxation time depth variation curve based on the final inversion model vector, and determines the top and bottom interface burial depth of the salt lake brine reservoir and corresponding physical parameters according to the gradient extreme position of the depth variation curve. High-resolution and high-stability inversion imaging of the water content and the transverse relaxation time of the salt lake brine reservoir is realized, and the top and bottom interface burial depth and the physical parameters of the brine reservoir can be accurately positioned.

Description

Technical Field

[0001] This application belongs to the field of geophysical exploration involving salt lake brine resources, specifically a magnetic resonance multi-scale focusing inversion method and system suitable for salt lake brine exploration. Background Technology

[0002] Salt lake brines are a major source of lithium and potassium minerals. Surface magnetic resonance (MRS) is currently the only geophysical method capable of directly and quantitatively detecting groundwater. Existing MRS inversion methods often employ smoothing constraints, resulting in good intra-layer continuity, but the top and bottom interfaces of aquifers are blurred, leading to large errors in the estimation of physical parameters and a bottleneck of insufficient inversion resolution. Focused inversion methods improve resolution by sharpening boundaries, but due to the acquisition characteristics of denser excitation pulse moments in shallower regions and sparser ones in deeper regions, the amount of data information decreases with increasing formation depth, making it difficult to adapt fixed focusing parameters to the scale differences at different depths. Furthermore, traditional focused inversion algorithms are highly dependent on the initial model, limiting convergence efficiency. Therefore, developing a high-resolution magnetic resonance inversion method that adaptively adjusts the focusing intensity with depth and introducing an efficient solution strategy is of great significance for high-precision evaluation of brine reservoirs.

[0003] In the field of magnetic resonance inversion, Chinese patent publication number CN121956174A discloses "an adaptive lateral constraint inversion method for nuclear magnetic resonance in complex geological conditions." This method constructs an objective function based on Tikhonov regularization and enhances the spatial continuity of the inversion model by fusing self-smoothing correlation coefficients and Pearson correlation coefficients. However, this method uses smoothing constraints, resulting in insufficient vertical resolution and difficulty in accurately identifying aquifer interfaces.

[0004] Regarding focused inversion, Chinese patent publication number CN122064907A discloses a "gravity adaptive focused inversion method, system, storage medium, and device based on predicted field migration and dynamic weights." This method couples an adaptive weight matrix and a focused constraint matrix to improve the depth resolution of the inversion. However, this method has a fixed focusing scale and cannot dynamically adjust the focusing intensity according to changes in data density at different depths, which limits its applicability in complex geological structures.

[0005] Regarding the solution method, Chinese Patent Publication No. CN117388935A discloses "A Three-Dimensional Focusing Inversion Method for Electrical Source-Ground Transient Electromagnetic Interfaces," which achieves high-resolution focusing imaging of electrical interfaces based on L1 norm regularization constraints. However, this method uses the conjugate gradient method, which is highly dependent on the initial model and requires multiple line searches and direction updates, thus limiting the inversion efficiency and stability. Summary of the Invention

[0006] This application provides a magnetic resonance multi-scale focusing inversion method and system suitable for salt lake brine detection, which solves the problems of low resolution, difficulty in adapting to different depth scale differences, and limited convergence efficiency in salt lake brine magnetic resonance inversion.

[0007] The first aspect of this application provides a magnetic resonance multi-scale focusing inversion method suitable for detecting brine in salt lakes, including: The observation magnetic resonance signal of the detection area is collected, the underground space of the detection area is divided into a one-dimensional layered medium model, the water content and transverse relaxation time of each discrete layer are used as the physical property parameters to be inverted, the inversion model vector is constructed, and a full waveform forward modeling operator is established to generate the predicted magnetic resonance signal corresponding to the current inversion model vector. An inversion objective function is constructed by superimposing an observation response functional and a joint regularization term. The observation response functional is used to measure the fitting residual between the predicted magnetic resonance signal and the observed magnetic resonance signal. The joint regularization term uses the difference in physical property parameters between adjacent layers in the inversion model vector as a gradient to apply a focusing constraint to the inversion model vector. The inversion objective function is transformed into an analytically differentiable approximation, and the momentum gradient descent method is used for iterative solution until the relative change in the model data fit difference satisfies the termination condition, and the final inversion model vector is output. Based on the final inversion model vector, the depth variation curves of water content and lateral relaxation time are extracted. The burial depth of the top and bottom interfaces of the salt lake brine reservoir and the corresponding physical parameters are determined according to the gradient extreme value positions of the depth variation curves.

[0008] Furthermore, the inversion objective function is: , In the formula, For the inversion objective function, To observe the response functional, For joint regularization terms, The inversion model vector is composed of water content. and lateral relaxation time Arrangement and composition, For full-waveform forward modeling analog operators, The difference in physical properties between adjacent strata, among which , , The number of strata. , For the first The water content of the first layer minus the first layer The moisture content of the layer, For the first The lateral relaxation time value of the layer minus the first Lateral relaxation time of the layer To observe magnetic resonance signals, Weighted matrix of data, These are the weighting coefficients of the joint normalization term. The weighting coefficients for the joint regularization term of water content are... The weight coefficients of the joint regularization term for the horizontal relaxation time; The focal scale factor for water content, The focal scale factor is the horizontal relaxation time.

[0009] Furthermore, by dividing the original signal into multiple time gates and taking the reciprocal of the noise amplitude within each time gate, a data weighting matrix is ​​constructed.

[0010] Furthermore, the focal scale factor of water content is in the th... The value at the nth iteration is equal to the water content focusing scale factor scaling factor multiplied by the nth iteration. The median of the absolute values ​​of the water content differences between all adjacent layers in the inversion model vector obtained from the next iteration; The focusing scale factor of the transverse relaxation time is in the first... The value at the nth iteration is equal to the scaling factor of the lateral relaxation time focus scale multiplied by the nth iteration. The median absolute value of the lateral relaxation time difference between all adjacent layers in the inversion model vector obtained from the next iteration.

[0011] Furthermore, a transition factor associated with the number of inversion iterations is set, and the inversion process is coordinated to be in coarse-scale focus or fine-scale focus based on the transition factor, and the weight coefficient of the joint regularization term is dynamically updated according to the transition factor.

[0012] Furthermore, the transition factor associated with the number of inversion iterations is defined as follows: , For the first Iteration transition factor This is the preset number of focus transition iterations; The weight coefficients of the joint regularization term, dynamically updated according to the transition factor, are expressed as follows: , In the formula, For the first The weight coefficients of the joint regularization term in the iteration. As the initial value, For the final value, when hour, The inversion is focused on a coarse scale. From initial value Increment to final value ;when hour, The inversion is focused on a fine scale. Fixed to final value .

[0013] Furthermore, the fine-scale focusing constructs a gradient adaptive penalty based on the stratigraphic profile obtained from coarse-scale focusing. The penalty intensity is dynamically adjusted according to the ratio of differences in physical properties between adjacent stratigraphic units to the focusing scale factor, and a relative threshold coefficient is set. and , When the ratio of the difference in physical properties between adjacent strata to the focusing scale factor is less than When the ratio is greater than 1, configure the first penalty weight; when the ratio is greater than 1, configure the first penalty weight. When the ratio is greater than or equal to the first penalty weight, a second penalty weight is configured, and the second penalty weight is greater than the first penalty weight; and less than or equal to When the penalty weight is linearly interpolated between the first penalty weight and the second penalty weight, the first penalty weight and the second penalty weight are used as the pre-weighting coefficients in the joint regularization term.

[0014] Furthermore, the inversion objective function is transformed into an analytically differentiable approximation, including: constructing a water content diagonal weight matrix and a lateral relaxation time diagonal weight matrix based on the gradient distribution of the inversion model vector, where the diagonal elements are the reciprocals of the sum of the squares of the differences in physical properties between corresponding adjacent layers and the squares of the focusing scale factor; based on the water content diagonal weight matrix and the lateral relaxation time diagonal weight matrix, approximating the joint regularization term as a weighted quadratic form of the physical property gradient vector, thus transforming the original non-convex and non-quadratic inversion objective function into an analytically differentiable approximation.

[0015] A second aspect of this application provides a magnetic resonance multi-scale focusing inversion system suitable for detecting brine in salt lakes, comprising: The data acquisition and modeling module is used to acquire the observed magnetic resonance signals in the detection area, divide the underground space of the detection area into a one-dimensional layered medium model, construct the inversion model vector with the water content and lateral relaxation time of each discrete layer as the physical property parameters to be inverted, and establish a full waveform forward modeling operator to generate the predicted magnetic resonance signal corresponding to the current inversion model vector. The objective function construction module is used to construct an inversion objective function composed of the superposition of the observation response functional and the joint regularization term. The observation response functional is used to measure the fitting residual between the predicted magnetic resonance signal and the observed magnetic resonance signal. The joint regularization term uses the difference in physical property parameters between adjacent layers in the inversion model vector as a gradient to apply a focusing constraint to the inversion model vector. The iterative inversion solution module is used to transform the inversion objective function into an analytically differentiable approximate form, and to perform iterative solution using the momentum gradient descent method until the relative change in the model data fit difference satisfies the termination condition, and outputs the final inversion model vector. The reservoir parameter extraction module is used to extract the depth variation curves of water content and lateral relaxation time based on the final inversion model vector, and to determine the burial depth of the top and bottom interfaces of the salt lake brine reservoir and the corresponding physical property parameters based on the gradient extreme value positions of the depth variation curves.

[0016] Compared with the prior art, this application has at least the following beneficial effects: This application achieves a balance between high-resolution layered interface identification and high-stability physical property parameter inversion in the magnetic resonance detection of salt lake brine by deeply integrating multi-scale focusing constraints and adaptive regularization strategies. It realizes high-resolution and high-stability inversion imaging of water content and lateral relaxation time of salt lake brine reservoirs, and can accurately locate the burial depth of the top and bottom interfaces of brine reservoirs and their physical property parameters. Attached Figure Description

[0017] Figure 1 A flowchart of a magnetic resonance multi-scale focusing inversion method for detecting salt lake brine provided in this application embodiment; Figure 2 The following are multi-scale focused inversion results provided in the embodiments of this application: (a) is the water content inversion curve, and (b) is the lateral relaxation time inversion curve. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] join Figure 1 As shown in the embodiment of this application, a magnetic resonance multi-scale focusing inversion method suitable for salt lake brine detection is provided, comprising: S101: Collect the observed magnetic resonance signal of the detection area, divide the underground space of the detection area into a one-dimensional layered medium model, use the water content and transverse relaxation time of each discrete layer as the physical property parameters to be inverted, construct the inversion model vector, and establish a full waveform forward modeling operator to generate the predicted magnetic resonance signal corresponding to the current inversion model vector. S102, construct an inversion objective function composed of the superposition of the observation response functional and the joint regularization term. The observation response functional is used to measure the fitting residual between the predicted magnetic resonance signal and the observed magnetic resonance signal. The joint regularization term uses the difference in physical property parameters between adjacent layers in the inversion model vector as a gradient to apply a focusing constraint to the inversion model vector. S103, the inversion objective function is transformed into an analytically differentiable approximation form, and the momentum gradient descent method is used for iterative solution until the relative change in the model data fitting difference satisfies the termination condition, and the final inversion model vector is output. S104. Based on the final inversion model vector, extract the depth variation curves of water content and lateral relaxation time, and determine the burial depth of the top and bottom interfaces of the salt lake brine reservoir and the corresponding physical property parameters according to the gradient extreme value position of the depth variation curves.

[0020] In step S101, a ground magnetic resonance detection system is deployed in the target area of ​​the salt lake brine. A multi-turn square or circular transmitting coil and receiving coil are used to transmit an alternating electromagnetic field with a specific frequency and pulse moment into the ground, which excites hydrogen protons in the groundwater to generate a macroscopic magnetic moment. After the transmission pulse ends, the raw signals of the free induction decay signal or spin echo signal are collected. The collected raw signals are preprocessed to obtain the observed magnetic resonance signal.

[0021] Preprocessing is performed on the acquired raw signals, including: Power frequency notch filtering: Adaptive notch filters or comb filters are used to eliminate 50Hz and its integer multiples of harmonic interference in order to suppress the coupling interference of industrial power grid to magnetic resonance signals. Random noise suppression: Singular value decomposition or wavelet transform threshold denoising method is used to decompose and reconstruct the acquired signal, suppress environmental random noise and instrument background noise, and improve the signal-to-noise ratio; Phase correction and zero-time alignment: Phase rotation or Hilbert transform methods are used to eliminate the phase offset between the transmitted pulse and the received signal, and all signals are aligned to the zero-time reference to ensure that the time reference reference between the observed data and the subsequent forward modeling prediction data is consistent. The underground space of the detection area is discretized into a one-dimensional layered medium model. Specifically, along the depth direction, the underground space is divided into equal or varying thickness sections from the surface downwards. Each layer is a horizontally layered unit, with a fixed thickness. The physical properties within each layer are considered uniformly distributed horizontally, changing only in the vertical direction. Total number of layers. Based on the depth and resolution requirements of the detection target, in one specific embodiment of this application, N is set to 35 layers.

[0022] In the one-dimensional layered medium model, the physical properties of each layer include water content and transverse relaxation time. Water content is the volume fraction of water in the porous medium of each layer, characterizing the water storage capacity of that layer; transverse relaxation time characterizes the attenuation characteristics of the transverse magnetization vector of the aqueous fluid in that layer under the influence of magnetic field inhomogeneity, and is an important physical property parameter for distinguishing the properties of fluids with different pore structures or different mineralization.

[0023] The water content and transverse relaxation time of each discrete layer in the above one-dimensional layered medium model are used as unknown physical property parameters to be inverted. These parameters are arranged in a fixed order to construct an inversion model vector. The inversion model vector is a vector with dimension... The vector is a one-dimensional array, with the first half representing the water content sequence of each layer and the second half representing the lateral relaxation time sequence of each layer. Each element in this inversion model vector is an unknown quantity to be inverted, and the goal of the inversion algorithm is to find the optimal estimate of each element in the inversion model vector through iterative optimization.

[0024] In this embodiment, the full-waveform forward modeling operator is established based on the Bloch equation or a modified Bloch equation to describe the physical mapping relationship between the subsurface medium properties (water content and transverse relaxation time) and the induced electromotive force in the ground receiving coil over time. The physical mechanism of magnetic resonance detection is as follows: In the detection of brine in salt lakes, an alternating current pulse with a frequency equal to the Larmor frequency is applied to the underground medium through a ground-based transmitting coil. Under the influence of the geomagnetic field, a transverse magnetization vector perpendicular to the direction of the geomagnetic field is generated. After the transmitting pulse ends, the transmitting current is cut off. During the process of the transverse magnetization vector of hydrogen protons in the groundwater returning to an equilibrium state, a free-induction decay signal that decays over time is generated in the receiving coil. The initial amplitude of this free-induction decay signal is proportional to the groundwater content, and the decay rate of the free-induction decay signal is related to the transverse relaxation time.

[0025] The full-waveform forward modeling operator is: given a set of discrete layer water content and transverse relaxation time parameters (i.e., the current inversion model vector), it simulates the complete waveform of the induced electromotive force in the receiving coil changing with time through numerical calculation.

[0026] The construction process of the full-waveform forward modeling operator is as follows: Kernel function calculation: Based on the geometric parameters of the transmitting and receiving coils, the transmitting pulse moment, the geomagnetic field strength and gradient, the conductivity distribution of the underground medium, and other parameters, the response kernel function of each layer to the magnetic resonance signal is calculated. This kernel function reflects the contribution coefficient of the unit water content of the layer to the amplitude of the received signal and its attenuation characteristics over time. Signal synthesis: The water content of each layer is multiplied by the corresponding kernel function, and the exponential decay caused by the transverse relaxation time is considered. The contributions of all layers are weighted and superimposed to generate a full-time magnetic resonance free induction decay signal or spin echo signal. Instrument response convolution: The theoretical signal is convolved with the transfer function of the instrument system (including the frequency response and phase response of the receiving filter) to simulate the filtering and distortion effects of the instrument on the theoretical signal, and generate a predicted magnetic resonance full waveform signal with the same sampling rate and time base as the measured signal.

[0027] The full-waveform forward modeling operator is denoted as Its input is the current inversion model vector, and its output is the predicted magnetic resonance signal corresponding to the model. In each iteration, the current inversion model vector is substituted into this full-waveform forward modeling operator to calculate the current predicted magnetic resonance signal, which is used for comparison with the observed magnetic resonance signal.

[0028] In one embodiment, step S102 constructs an inversion objective function consisting of the superposition of the observation response functional and the joint regularization term. The inversion objective function is composed of the superposition of the observation response functional and the joint regularization term, and its basic form is as follows: , in, For the inversion model vector, Observational response functional For joint regularization terms, These are the weighting coefficients of the joint regularization term, used to adjust the overall strength of the regularization constraint in the inversion objective function. The objective function for inversion is defined as follows: The observation response functional ensures that the inversion results do not deviate from the true information contained in the measured data, while the joint regularization term imposes prior geological constraints to overcome the ambiguity of the inversion. The two are balanced between data-driven and prior constraints through the weight coefficient of the joint regularization term.

[0029] Observational response functional It is used to quantitatively describe the fit between the predicted magnetic resonance signal generated by the full waveform forward modeling operator and the actual acquired observed magnetic resonance signal. It is defined as the square of the second norm of the difference between the predicted magnetic resonance signal and the observed magnetic resonance signal.

[0030] , To observe magnetic resonance signals, Weighted matrix for the data.

[0031] The core function of the joint regularization term is to apply a focusing constraint to the inversion model vector by using the difference in physical property parameters between adjacent layers in the inversion model vector as a gradient. The so-called focusing constraint refers to using a specific mathematical penalty function to make the inversion results exhibit a stepped distribution with uniform blocks and abrupt interface changes in the vertical depth profile, thereby highlighting the physical property abrupt boundary between the brine reservoir and its surrounding rock, and achieving high-resolution identification of the top and bottom interfaces of the reservoir.

[0032] The physical basis of this constraint stems from the geological structural characteristics of brine reservoirs in salt lakes: in salt lake sedimentary environments, brine reservoirs are typically situated within specific sedimentary cycles or lithological intervals, with relatively uniform water content and pore structure within the reservoir, while significant abrupt changes in physical properties exist between the reservoir and the surrounding rocks (such as mudstone, salt rock layers, or dense carbonate layers). Therefore, the true vertical distribution of physical parameters should exhibit a structural characteristic of smooth intra-layer distribution and abrupt inter-layer distribution. The focusing constraint is precisely the mathematical expression of this geological a priori information. By penalizing the number of non-zero elements in the gradient distribution, it ensures that the inversion results retain only a few significant physical property abrupt transition interfaces necessary for fitting the observed data.

[0033] Traditional smoothing constraints impose a penalty on all gradient components with similar strengths, which weakens the gradient at the real interface and produces a blurred transition band. In contrast, focusing constraints differentiate the gradient components, compressing very small gradient components to near zero (intra-layer homogenization) and preserving or even enhancing very large gradient components (interface sharpening), thereby achieving sparsity focusing of the physical property profile.

[0034] In this embodiment, the focusing constraints of water content and lateral relaxation time are combined to form a joint regularization term, the complete form of which is: , in, For joint regularization terms, The inversion model vector is composed of water content. and lateral relaxation time Arrangement and composition, For full-waveform forward modeling analog operators, The difference in physical properties between adjacent strata, among which , , The number of strata. , For the first The water content of the first layer minus the first layer The moisture content of the layer, For the first The lateral relaxation time value of the layer minus the first Lateral relaxation time of the layer The weighting coefficients for the joint regularization term of water content are... The weight coefficients of the joint regularization term for the horizontal relaxation time; The focal scale factor for water content, The focal scale factor is the horizontal relaxation time.

[0035] For water content focusing constraints, This is a lateral relaxation time focusing constraint.

[0036] The physical mechanism of the synergistic effect between the water content focusing constraint and the transverse relaxation time focusing constraint is as follows: Water content focusing constraints: These directly constrain the vertical distribution pattern of water content, resulting in a clear layered structure in the inverted water content profile. Water content is the most direct indicator parameter of brine reservoirs in salt lakes, and abrupt water content changes typically correspond to the top and bottom boundaries of the reservoir.

[0037] Lateral relaxation time focusing constraints: These indirectly constrain the longitudinal variations in pore structure and fluid properties, causing the lateral relaxation time profile to also exhibit a layered structure. Lateral relaxation time is sensitive to changes in pore structure, salinity, and hydrogen content; its abrupt changes at the interface can serve as independent verification information for water content interface identification.

[0038] The advantage of combining the two constraints is that water content and lateral relaxation time have different response mechanisms to the geological interface (water content mainly reflects changes in water storage capacity, while relaxation time mainly reflects changes in pore size and fluid properties). Both parameters undergo significant gradient changes at the same interface. By combining the focusing constraints, the interface positions of the two parameters tend to be consistent, which further enhances the geological reliability of the interface identification results and effectively reduces the multivariate problem of single parameter inversion.

[0039] After applying a focusing constraint through a joint regularization term, the longitudinal physical property profile of the inversion result will exhibit the following characteristics: Within the brine reservoir, the water content and the difference in lateral relaxation time between adjacent layers are compressed to near zero, forming a layer with uniform physical properties. At the interface between the reservoir and the surrounding rock, water content and lateral relaxation time undergo significant abrupt changes, and the gradient amplitude is preserved, forming a clear step-like jump. Within the non-reservoir surrounding rock section, physical properties remain relatively constant, and the gradient is also compressed to near zero.

[0040] Therefore, the focus constraint makes the entire inversion profile mathematically equivalent to an approximation result of a piecewise constant function, where the location and number of segments are determined by both data fitting requirements and prior geological information. This piecewise constant characteristic closely matches the actual geological structure of the brine reservoir, thus ensuring the geological rationality of the inversion results from a physical mechanism perspective.

[0041] In one embodiment, the data weighting matrix in the inversion objective function is obtained by dividing the original signal into multiple time gates and taking the reciprocal of the noise amplitude within each time gate.

[0042] In magnetic resonance signal acquisition, signals at different time points exhibit significantly different signal-to-noise ratio (SNR) characteristics. The amplitude of the magnetic resonance free-induction decay signal is largest in the early period after the transmission pulse ends, with concentrated signal energy and the highest SNR. As time progresses, the signal decays exponentially, and the amplitude gradually decreases, while the ambient noise level remains relatively constant or changes slowly, leading to a sharp drop in the SNR of the later signals. If all time sampling points are assigned the same weight during the inversion process, the low SNR data from the later stages will introduce a large relative error into the fitting residuals, affecting the stability and reliability of the inversion results. To address this issue, this embodiment divides the original signal into multiple time gates and constructs a data weighting matrix by taking the reciprocal of the noise amplitude within each time gate.

[0043] Before each magnetic resonance emission pulse is applied, the system is in receiving mode but without emission pulse excitation. The signal acquired during this time consists of pure ambient noise and instrument background noise. The acquisition period is [duration duration missing]. The baseline noise signal is divided into multiple intervals using the same time-gating method as the magnetic resonance signal. The root mean square value of the noise samples within each interval is calculated as the noise amplitude estimate for that time gate. The reciprocal of the noise amplitude is then used to construct a data weighting matrix. This data weighting matrix is ​​a diagonal matrix with dimensions of [dimensions missing]. (Total number of data sampling points), of which The number of pulse moments. The number of time sampling points for each pulse moment signal. For the th Under the condition of the nth pulse moment, the first The weighting coefficients corresponding to each time sampling point are determined according to the time gate to which it belongs.

[0044] In some embodiments, the values ​​of the two focusing scale factors in the regularization term have a decisive influence on the focusing effect of the regularization term.

[0045] The focal scale factor of water content is in the first place. The value at the nth iteration is equal to the water content focusing scale factor scaling factor multiplied by the nth iteration. The median of the absolute values ​​of the water content differences between all adjacent layers in the inversion model vector obtained from the next iteration; The focusing scale factor of the transverse relaxation time is in the first... The value at the nth iteration is equal to the scaling factor of the lateral relaxation time focus scale multiplied by the nth iteration. The median absolute value of the lateral relaxation time difference between all adjacent layers in the inversion model vector obtained from the next iteration.

[0046] Represented as: ; , In the formula, , , respectively, are the focal scale factor scaling factors for water content and the focal scale factor scaling factors for transverse relaxation time. For the number of iterations, The focal scale factor for water content, This is the focusing scale factor for the transverse relaxation time. For the first The inversion model vector obtained in the nth iteration The water content of the first layer minus the first layer The moisture content of the layer, For the first The inversion model vector obtained in the nth iteration The lateral relaxation time value of the layer minus the first The lateral relaxation time value of the layer.

[0047] Different scaling factors are used for water content and transverse relaxation time because the gradient distribution characteristics of these two physical properties are fundamentally different. Water content ranges from 0 to 1 (volume fraction), and its gradient magnitude is correspondingly constrained within this range, resulting in a relatively compact distribution. A relatively small scaling factor is sufficient for effective differentiation. The numerical range of the transverse relaxation time depends on specific geological conditions, and can vary from a few milliseconds to hundreds of milliseconds. The gradient distribution range is much larger than that of water content, and the distribution pattern may be more dispersed. Therefore, a smaller scaling factor is used. A larger scaling factor is used to accommodate its wider gradient distribution range, ensuring that the focus scale factor can cover and distinguish the relative differences between gradients of different magnitudes.

[0048] It should be noted that, and This is a preferred value in one specific embodiment of this application. In practical applications, it can be appropriately adjusted according to the magnitude and variation characteristics of the physical property parameters in different work areas. However, regardless of the value of the scaling factor, it should be greater than 1 to ensure that the focusing scale factor is slightly higher than the median level.

[0049] In some embodiments, a transition factor associated with the number of inversion iterations is set, and the inversion process is coordinated to be in coarse-scale focus or fine-scale focus based on the transition factor, and the weight coefficient of the joint regularization term is dynamically updated according to the transition factor.

[0050] In magnetic resonance inversion, the timing and intensity of regularization constraints have a decisive impact on the convergence and final accuracy of the inversion results. If strong focusing constraints are applied at the beginning of the inversion, the initial model deviates significantly from the actual subsurface medium distribution. Strong constraints can easily lead the inversion process away from the local optimum of the actual data response, resulting in inaccurate fitting of the inversion results to the observed data. Conversely, if weak focusing constraints are maintained throughout the inversion process, although the data fitting accuracy is guaranteed, the inversion results lack prior guidance on layered structures, have insufficient interface resolution, and are difficult to accurately identify the top and bottom boundaries of brine reservoirs.

[0051] More complexly, the scale of the focusing constraints should dynamically change with the inversion process. In the early stages of inversion, the model parameters deviate significantly from the true values. At this point, coarse-scale focusing is needed, which constrains the macroscopic layered structure of the model on a large scale, allowing the inversion results to quickly converge to the correct layered framework, without being overly concerned with precise intralayer details. In the later stages of inversion, the model has roughly approximated the true distribution. At this point, fine-scale focusing is needed, which refines the interface locations at a local scale, sharpening the boundaries of the physical property profiles to the highest resolution. Coarse-scale and fine-scale focusing have drastically different requirements for the weighting coefficients of the regularization term and the gradient discrimination threshold; a single, fixed constraint strategy cannot satisfy both needs.

[0052] To address the aforementioned technical challenges, this application proposes a multi-scale focusing collaborative scheduling mechanism based on a transition factor. This mechanism establishes a transition factor associated with the number of inversion iterations, enabling a smooth transition between coarse-scale and fine-scale focusing during the inversion process. Simultaneously, it dynamically updates the weight coefficients of the joint regularization term based on changes in the transition factor, thereby achieving coordinated matching of focusing scale, focusing intensity, and iteration process.

[0053] The transition factor associated with the number of inversion iterations is defined as follows: , For the first Iteration transition factor This is the preset number of focus transition iterations; The weight coefficients of the joint regularization term, dynamically updated according to the transition factor, are expressed as follows: , In the formula, For the first The weight coefficients of the joint regularization term in the iteration. As the initial value, For the final value, when hour, The inversion is focused on a coarse scale. From initial value Increment to final value ;when hour, The inversion is focused on a fine scale. Fixed to final value .

[0054] The transition factor is a monotonically non-decreasing parameter that takes values ​​in a closed interval from 0 to 1, and its value is determined entirely by the current inversion iteration number. The decision is independent of the state of the inversion model itself. The transition factor divides the entire inversion iteration process into two intervals: a coarse-scale focusing stage and a fine-scale focusing stage. The value of the transition factor reflects the relative position of the current iteration within these two intervals.

[0055] In the initial stage of the inversion iteration (when k=1), the transition factor is 0, indicating that it is completely under the control of coarse-scale focusing. As the number of iterations increases... The transition factor gradually increases linearly from 0, with larger values ​​indicating a move away from the coarse-scale focusing stage and closer to the fine-scale focusing stage. When the number of iterations reaches or exceeds the preset number of focusing transition iterations, the transition factor saturates to 1 and remains constant, indicating that it has now fully entered the fine-scale focusing stage and will no longer revert to the coarse-scale focusing mode.

[0056] The monotonically increasing and saturating nature of the transition factor ensures that the switch between coarse-scale and fine-scale focusing is unidirectional and smooth, rather than abrupt. This characteristic is crucial for the numerical stability of the inversion; if the two focuses switch abruptly, the shape of the inversion objective function will change drastically, potentially leading to oscillations or even divergence in the iterative process. The linearly increasing nature of the transition factor allows the relative roles of the two focuses to gradually increase and decrease over dozens of iterations, achieving a smooth transition between the two constraint mechanisms.

[0057] Coarse-scale focusing and fine-scale focusing are not two independent algorithms or two independent programs, but rather two constraint modes embodied by different configurations of weight coefficients and focusing scale factors in the joint regularization term within the same inversion framework.

[0058] Coarse-scale focusing corresponds to the early stage of inversion (i.e. (the interval), where the weight coefficient of the joint regularization term is... With smaller values, the regularization constraint has a lower overall weight in the inversion objective function. Under coarse-scale focusing, the focusing constraint focuses on constraining the macroscopic layered structure within a large-scale range in the depth domain. Specifically, it manifests as a global convergence constraint on multiple adjacent discrete layers (e.g., more than 5 layers), prompting the inversion model vector to form a block-uniform layered framework at the macroscopic level, rather than pursuing precise sharpening of the interface position of each layer. The goal of the coarse-scale focusing stage is to enable the inversion to quickly converge to a large-scale structure that is basically consistent with the observed data, laying the foundation for subsequent fine-scale focusing. In this stage, due to the weak regularization constraint and the coarse scale of action, the inversion algorithm has sufficient degrees of freedom to search within a wide region of the data space, making it less prone to getting trapped in local optima.

[0059] Fine-scale focusing corresponds to the later stages of the inversion (i.e. (the interval), where the weight coefficient of the joint regularization term is... It has increased to the final value. And it remains constant, with regularization constraints at their strongest level. In the fine-scale focusing mode, the focusing constraint focuses on refining and sharpening the differences in physical properties between adjacent single layers. Specifically, it involves independently evaluating and penalizing the gradient components of each pair of adjacent layers, so that the gradient magnitude at the true interface is preserved, while the gradient magnitude at non-interfaces is compressed to near zero. The goal of the fine-scale focusing stage is to refine the precise location of each interface based on the established layered framework, so that the physical property profile achieves the highest longitudinal resolution.

[0060] The dynamic updating of the weight coefficients of the joint regularization term is achieved under the drive of the transition factor. This ensures that the strength of the regularization constraint evolves synchronously with the progress of the inversion process. Specifically, weaker constraints are applied in the early stages of the inversion to ensure the dominance of data fitting and global search capability, while constraints are gradually strengthened in the later stages of the inversion to highlight the focusing effect and interface sharpening capability, thereby achieving a dynamic balance between data fitting accuracy and model structure clarity.

[0061] Fine-scale focusing constructs an adaptive gradient penalty based on the stratigraphic profile obtained from coarse-scale focusing. It dynamically adjusts the penalty intensity according to the ratio of differences in physical properties between adjacent strata to the focusing scale factor, and sets a relative threshold coefficient. and , When the ratio of the difference in physical properties between adjacent strata to the focusing scale factor is less than When the ratio is greater than 1, configure the first penalty weight; when the ratio is greater than 1, configure the first penalty weight. When the ratio is greater than or equal to the first penalty weight, a second penalty weight is configured, and the second penalty weight is greater than the first penalty weight; and less than or equal to When the penalty weight is linearly interpolated between the first penalty weight and the second penalty weight, the first penalty weight and the second penalty weight are used as the pre-weighting coefficients in the joint regularization term; That is, the joint regularization term is rewritten as: , in, This is the first pre-weighting coefficient. The second pre-weighting coefficient takes the value of the first penalty weight, the second penalty weight, or a linear interpolation between the first penalty weight and the second penalty weight.

[0062] In some embodiments, the inversion objective function is transformed into an analytically differentiable approximation, including: constructing a water content diagonal weight matrix and a lateral relaxation time diagonal weight matrix based on the gradient distribution of the inversion model vector, wherein the diagonal elements are the reciprocals of the sum of the squares of the differences in physical properties between corresponding adjacent layers and the squares of the focusing scale factor; based on the constructed water content diagonal weight matrix and lateral relaxation time diagonal weight matrix, approximating the joint regularization term as a weighted quadratic form of the physical property gradient vector, so that the original non-convex and non-quadratic inversion objective function is transformed into an analytically differentiable approximation.

[0063] Specifically, this includes: the process of transforming the inversion objective function into an approximate form, based on the iterative reweighting framework, in the... In the next iteration, a water content diagonal weight matrix is ​​constructed based on the gradient distribution of the current inversion model vector. and the diagonal weight matrix of the lateral relaxation time : ; ; Combine regularization terms It is approximately a weighted quadratic form: , Here, the property gradient vector refers to a one-dimensional array composed of the differences in property parameters between adjacent layers after the first-order difference transformation of the property parameters of each discrete layer in the inversion model vector. and The property gradient vector characterizes the rate and magnitude of change of two key physical parameters (water content and lateral relaxation time) in the underground medium along the vertical depth direction.

[0064] This leads to the inversion objective function, which is neither convex nor quadratic. Transform into an approximate form of the inverse objective function that can be analytically differentiated. : ; In the formula, It is a first-order difference matrix. For the first The diagonal weight matrix of the joint regularization term in the next iteration . For the first In the nth iteration The diagonal matrix elements corresponding to each pair of adjacent water content layers For the first In the nth iteration The diagonal matrix elements corresponding to adjacent layer pairs with lateral relaxation times.

[0065] The momentum gradient descent method is used to iteratively solve the inversion objective function until the relative change in the model data fit difference is less than 1%, and the final inversion model vector is output.

[0066] After the inversion objective function has been transformed into an analytically differentiable approximation, an efficient and stable optimization algorithm is needed to find the optimal inversion model vector that minimizes the inversion objective function. This application selects momentum gradient descent as the core algorithm for iterative solution, based on the following considerations: Traditional gradient descent only uses the current gradient information to determine the model update direction in each iteration. In high-dimensional model spaces and complex objective function forms, it is prone to getting stuck in local extrema or oscillations, resulting in slow convergence. Momentum gradient descent introduces a momentum term, incorporating historical gradient information into the current update direction in a decaying and accumulating manner. This makes the model update not only depend on the gradient of the current objective function but also on the continuous influence of the previous update direction and magnitude. This mechanism has significant advantages in inversion problems. It can accelerate traversal in relatively flat regions of the inversion objective function using accumulated momentum and suppress oscillations in regions where the gradient direction of the objective function changes frequently using momentum inertia, thereby achieving faster and more stable convergence.

[0067] The momentum gradient descent method is implemented as follows: In each iteration, the gradient of the inversion objective function in the current approximate form with respect to the inversion model vector is first calculated, i.e., the upward direction of the inversion objective function in the model space; then, the momentum term from the previous iteration is combined with a weighted sum to obtain the update direction for this iteration; finally, the inversion model vector is corrected along this update direction according to a preset learning rate to form a new iterative model. This process is repeated until the convergence condition is met.

[0068] In each iteration, the momentum term is equivalent to a decay-weighted average of all historical gradient information, with the gradient from the most recent iteration contributing the most, while the gradient contributions from earlier iterations gradually decrease according to an exponential decay law. This design gives the momentum term two complementary adjustment functions.

[0069] The first function is the acceleration effect. When the inversion objective function resembles a long and narrow canyon in the model space, the traditional gradient descent method oscillates repeatedly between the canyon walls, advancing only a tiny distance along the canyon's direction each time, resulting in extremely slow convergence. The momentum term, by accumulating the gradient components along the canyon's direction, continuously accelerates the model update direction along the canyon's direction, significantly improving the speed of advancement in flat or gently sloping regions and reducing the number of iterations required to reach convergence.

[0070] The second function is the damping effect. When the gradient direction of the objective function frequently reverses in adjacent iterations (e.g., oscillating back and forth between the walls of a narrow canyon), the historical gradient direction in the momentum term partially cancels out the current gradient direction, effectively suppressing the oscillating component perpendicular to the canyon's direction. This makes the model update path smoother and improves the stability of the iteration process. The synergistic effect of these two effects enables the momentum gradient descent method to efficiently and stably search for the optimal solution in the solution space of the inversion problem.

[0071] The complete process of solving the inversion objective function using the momentum gradient descent method begins with the initialization phase. Before the inversion begins, initial guess values ​​need to be assigned to the inversion model vectors, usually based on prior geological information of the work area or the assumption of a uniform half-space. Simultaneously, the initial value of the momentum term is set to a zero vector, indicating that no historical gradient information has been accumulated during the first iteration.

[0072] In each iteration, a forward modeling calculation is first performed on the current inversion model vector. This involves inputting the current inversion model vector into a full-waveform forward modeling operator to generate the corresponding predicted magnetic resonance signal, and then calculating the fitting residual between this signal and the observed signal. Next, based on the analytical differentiability of the inversion objective function, the gradient of the inversion objective function at the current inversion model vector is calculated. This gradient reflects the sensitivity of the inversion objective function to each component of the model parameters at the current model position. Based on the direction and magnitude of the gradient, and combined with historical information from the momentum term in the previous step, the update direction for this iteration is calculated: the current gradient and the momentum term from the previous iteration are each superimposed in a certain proportion to form a new momentum term, which represents the direction of the current inversion model vector update.

[0073] After determining the update direction, the inversion model vector is corrected along that direction using the learning rate as the step size. The learning rate is a pre-set positive number that controls the magnitude of the update in each iteration. If the learning rate is too large, the model may overshoot the optimal solution and oscillate back and forth on both sides of the solution space or even diverge; if the learning rate is too small, the convergence speed is too slow, requiring a large number of iterations to achieve an acceptable result. In the preferred embodiment of this application, the value of the learning rate is dynamically determined during the iteration process through an adaptive line search strategy to ensure that each update allows the inversion objective function to decrease sufficiently without becoming excessive.

[0074] After the update is completed, one iteration is finished. Repeat the above process, and each iteration generates a new inversion model vector. The model cost of this inversion model vector in the sense of the inversion objective function decreases successively, the fit between the predicted magnetic resonance signal and the observed magnetic resonance signal increases successively, and the layered structure characteristics of the model gradually emerge and are continuously refined.

[0075] The termination of the iterative solution is controlled by the convergence criterion. The convergence criterion used in this application is: the iteration terminates when the relative change in the model data fit difference is less than a preset threshold.

[0076] The model data fit difference is a quantitative measure of the fitting residual between the predicted and observed magnetic resonance signals, and its specific value is calculated by the observation response functional. After each iteration, the fit difference between the predicted and observed magnetic resonance signals corresponding to the current model is calculated and compared with the fit difference of the previous iteration, and the relative change between the two is calculated. This relative change reflects the improvement in data fitting accuracy between two adjacent iterations.

[0077] When the relative change is less than a preset threshold, it indicates that continued iteration can no longer significantly improve the data fitting accuracy, the inversion process has entered the convergence plateau region, the model parameters tend to stabilize, and further iterations can only adjust the model within a very small range without making a substantial contribution to the data fitting. Terminating the iteration at this point is reasonable, as it ensures the accuracy of the inversion results while avoiding unnecessary computational overhead.

[0078] When the convergence condition is met, the iterative loop terminates, and the current inversion model vector is output as the final inversion result. This final vector contains the water content and lateral relaxation time values ​​of each discrete layer in the one-dimensional layered medium model, representing the optimal estimate of the distribution of real subsurface physical parameters by the inversion algorithm. This result fulfills two requirements simultaneously: in terms of data fitting, the fitting residual between the predicted magnetic resonance signal and the measured observation signal reaches a minimum, ensuring that the inversion result does not deviate from the true information contained in the measured data; in terms of structural constraints, through successive refinement by coarse-scale focusing and fine-scale focusing, the result exhibits clear layered structural characteristics, with homogeneous intralayer properties and sharp interlayer interfaces, providing a high-quality data foundation for the accurate extraction of subsequent reservoir parameters.

[0079] Step S104: Extract the depth variation curves of water content and lateral relaxation time based on the final inversion model vector, and determine the burial depth of the top and bottom interfaces of the salt lake brine reservoir and the corresponding physical property parameters based on the gradient extreme value positions of the depth variation curves.

[0080] Once the iterative convergence condition is met, the inversion process terminates, outputting the final inversion model vector. This vector consists of two sequences of physical property parameters: the first half represents the water content values ​​of each discrete layer, arranged in ascending order of depth; the second half represents the transverse relaxation time values ​​of each discrete layer, also arranged in ascending order of depth. Each discrete layer has its corresponding depth range, the total number of layers is determined before the inversion begins, and the top and bottom interface depths of each layer are fixed during model initialization. Therefore, the final inversion model vector actually contains two curves of physical property parameters that vary with depth: the water content curve as a function of depth and the transverse relaxation time curve as a function of depth.

[0081] The water content variation curve with depth reflects the vertical distribution characteristics of water content in the underground medium. In the exploration of brine in salt lakes, the brine reservoir typically exhibits a region of significantly high water content, while the surrounding rocks (such as mudstone layers, salt rock layers, or dense carbonate layers) exhibit regions of low water content. Therefore, the location where the water content curve abruptly changes from low to high indicates the top interface of the reservoir, and the location where it abruptly changes from high to low indicates the bottom interface of the reservoir.

[0082] The transverse relaxation time curve as a function of depth reflects the vertical distribution characteristics of pore structure and fluid properties in the subsurface medium. In brine reservoirs of salt lakes, because brine is usually hosted in sandstone or carbonate rocks with well-developed pores, its pore size and fluid properties differ from those of the surrounding rocks, resulting in different numerical characteristics of transverse relaxation time in the reservoir section and the surrounding rock section. Generally, due to high water content and good pore connectivity, the reservoir section has a relatively longer transverse relaxation time or exhibits a specific distribution pattern, contrasting with the surrounding rocks. Therefore, the abrupt change location of the transverse relaxation time curve can also serve as an independent criterion for interface identification.

[0083] The gradient of a depth variation curve refers to the rate of change of physical properties with depth, which is the difference in physical properties between adjacent layers divided by the interlayer spacing. In a discrete model with equal thickness, since all layers have the same thickness, the gradient simplifies to the difference in physical properties between adjacent layers. The magnitude of the gradient value directly reflects the degree of drastic change in physical properties in the longitudinal direction: when the physical property changes gradually within a certain depth range, the gradient value is close to zero; when a sudden change occurs in physical properties at a certain depth location, the gradient value exhibits an extreme value (maximum or minimum, depending on the direction of the abrupt change).

[0084] The gradient of the water cut curve exhibits a positive extreme value at the top interface of the reservoir because of the abrupt change from low water cut in the surrounding rock to high water cut in the reservoir; and a negative extreme value at the bottom interface of the reservoir because of the abrupt change from high water cut in the reservoir to low water cut in the surrounding rock, where the difference is negative and has a large absolute value. Therefore, the maximum value of the first derivative of the water cut curve corresponds to the top interface of the reservoir, and the minimum value corresponds to the bottom interface of the reservoir.

[0085] The gradient characteristics of the transverse relaxation time curve depend on the relative magnitudes of the relaxation times between the reservoir and the surrounding rock. If the reservoir's transverse relaxation time is longer than that of the surrounding rock, the gradient extremum follows a pattern similar to that of water cut, with positive extrema corresponding to the top interface and negative extrema corresponding to the bottom interface. If the reservoir's transverse relaxation time is shorter than that of the surrounding rock, the pattern is reversed. In either case, the location with the largest absolute gradient value always corresponds to the location of the most drastic change in physical properties, i.e., the geological interface.

[0086] Based on the above gradient analysis principle, this application uses an automated extreme value detection method to determine the burial depth of the top and bottom interfaces of the salt lake brine reservoir. The specific steps are as follows: The numerical calculation of the gradient curves involves calculating the difference between adjacent layers for the water content sequence and the lateral relaxation time sequence in the final inversion model vector, respectively, to obtain the water content gradient curve and the lateral relaxation time gradient curve. The horizontal axis of these two gradient curves represents the layer number (corresponding depth), and the vertical axis represents the gradient value. In practice, to eliminate the interference of small fluctuations within the layer on extreme value detection, the gradient curves can first be processed by median filtering or Gaussian smoothing to make the determination of extreme value locations more robust.

[0087] The automatic search for extreme value locations involves searching the water cut gradient curve from shallow to deep layers, identifying locations where the gradient value is greater than a certain positive threshold and is a local maximum. The corresponding depths are recorded as candidate locations for the top interface of the reservoir. The search continues deeper, identifying locations where the gradient value is less than a certain negative threshold and is a local minimum. The corresponding depths are recorded as candidate locations for the bottom interface of the reservoir. Similarly, the same search is performed on the lateral relaxation time gradient curve to obtain another set of candidate locations for the top and bottom interfaces. If the abrupt change in reservoir parameters is opposite to that of water cut, the rules for determining positive and negative extreme values ​​are adjusted accordingly.

[0088] The determination of the interface location using a multi-curve comprehensive approach is crucial. Since the water content curve and the transverse relaxation time curve reflect reservoir characteristics through different physical mechanisms, the interface location determined by these two curves may exhibit slight deviations due to data noise or geological complexity. When the difference between the extreme values ​​of the two curves is less than a preset tolerance range (e.g., the thickness of two layers), the weighted average of the two curves can be used as the final interface depth. The weighting coefficients can be allocated based on the relative sensitivity of the two curves to brine response. Typically, the water content curve has a greater weight than the transverse relaxation time curve because water content is the most direct indicator for reservoir identification. When the difference between the extreme values ​​of the two curves exceeds the tolerance range, it indicates a more complex geological structure or a lower signal-to-noise ratio for one curve. In this case, the following principles can be used for selection or combination: Prioritize the extreme value location of the water content curve, as water content provides the most direct indication of brine reservoirs; simultaneously, cross-validation can be performed using the morphology of the transverse relaxation time curve. If both curves exhibit gradient anomalies within the same depth range, even with slight deviations in the extreme value locations, the interface depth range can be comprehensively determined.

[0089] Thickness constraints and geological rationality verification: To prevent misjudgments caused by noise, after determining the top and bottom interfaces, the reservoir thickness needs to be calculated and compared with prior geological information for the work area (such as the known reasonable range of reservoir thickness). If the calculated thickness far exceeds the prior range or is lower than the minimum resolvable thickness (usually 1 to 2 layers), the extreme value detection results should be re-examined. This may be due to misjudging intra-layer interference as an interface or ignoring weak interfaces. This step provides a reasonable tolerance handling mechanism to ensure the geological rationality of the interface identification results.

[0090] After determining the burial depths of the top and bottom interfaces of the reservoir, the physical property parameters at the corresponding depths are extracted from the final inversion model vector. Specifically, the layer number of the discrete layer containing the top interface is determined based on its burial depth, and the water content and lateral relaxation time of that layer are read as the physical property parameters at the top interface. Similarly, the layer number of the bottom interface is determined based on its burial depth, and the water content and lateral relaxation time of that layer are read as the physical property parameters at the bottom interface.

[0091] In addition to the physical properties at the top and bottom interfaces, statistical characteristics of the internal physical properties of the reservoir can also be extracted. All discrete layers between the top and bottom interfaces are considered as reservoir segments, and the mean, median, maximum, and minimum values ​​of water cut and lateral relaxation time within these segments are calculated to comprehensively describe the reservoir's physical properties. The mean and median reflect the overall water abundance and pore structure characteristics of the reservoir, while the maximum and minimum values ​​reflect the degree of heterogeneity within the reservoir. These parameters have direct guiding significance for calculating the amount of brine resources in salt lakes and evaluating reservoir quality.

[0092] Thus, this application's embodiments form a complete technical loop, from the acquisition and preprocessing of raw magnetic resonance signals, the establishment of a one-dimensional layered model, and the construction of a full-waveform forward modeling operator; to the establishment of an inversion objective function including focusing constraints; to the collaborative scheduling of multi-scale focusing modules based on transition factors; to the iterative reweighting framework transformation and momentum gradient descent method solution; and finally to the automatic identification of reservoir top and bottom interfaces and physical parameters based on the gradient extrema of depth variation curves. Each step of this method serves the ultimate goal: to provide high-resolution, high-reliability, and highly automated quantitative inversion results of reservoir parameters for the exploration of salt lake brine resources.

[0093] On the other hand, this application provides a magnetic resonance multi-scale focusing inversion system suitable for salt lake brine detection. The above-mentioned magnetic resonance multi-scale focusing inversion method suitable for salt lake brine detection can be interpreted accordingly. The data acquisition and modeling module is used to acquire the observed magnetic resonance signals in the detection area, divide the underground space of the detection area into a one-dimensional layered medium model, construct the inversion model vector with the water content and transverse relaxation time of each discrete layer as the physical property parameters to be inverted, and establish a full waveform forward modeling operator to generate the predicted magnetic resonance signal corresponding to the current inversion model vector. The objective function construction module is used to construct an inversion objective function composed of the superposition of the observation response functional and the joint regularization term. The observation response functional is used to measure the fitting residual between the predicted magnetic resonance signal and the observed magnetic resonance signal. The joint regularization term uses the difference in physical property parameters between adjacent layers in the inversion model vector as a gradient to apply a focusing constraint to the inversion model vector. The iterative inversion solution module is used to transform the inversion objective function into an analytically differentiable approximate form, and to perform iterative solution using the momentum gradient descent method until the relative change in the model data fit difference satisfies the termination condition, and outputs the final inversion model vector. The reservoir parameter extraction module is used to extract the depth variation curves of water content and lateral relaxation time based on the final inversion model vector, and to determine the burial depth of the top and bottom interfaces of the salt lake brine reservoir and the corresponding physical property parameters based on the gradient extreme value positions of the depth variation curves.

[0094] It also includes a multi-scale focusing control module, which is used to set a transition factor associated with the number of inversion iterations, coordinate the inversion process to be in the coarse-scale focusing stage or the fine-scale focusing stage according to the transition factor, and dynamically update the weight coefficient of the joint regularization term according to the transition factor. In the multi-scale focusing control module, the transition factor is a monotonically non-decreasing function. The transition factor is less than 1 when the number of inversion iterations is less than the preset number of focusing transition iterations, and equal to 1 and remains constant when the number of inversion iterations is greater than or equal to the preset number of focusing transition iterations. The weight coefficient of the joint regularization term increases linearly with the transition factor. When the transition factor is 0, the initial weight coefficient is taken, and when the transition factor is 1, the final weight coefficient is taken.

[0095] In some embodiments, the system further includes a focus scale adaptive update module, used to update the focus scale factor of water cut and the focus scale factor of lateral relaxation time respectively in each iteration based on the gradient statistical distribution of the current inversion model vector; the focus scale factor of water cut is equal to the water cut scaling factor multiplied by the median absolute value of the water cut difference between adjacent layers; the focus scale factor of lateral relaxation time is equal to the lateral relaxation time scaling factor multiplied by the median absolute value of the lateral relaxation time difference between adjacent layers; the water cut scaling factor and the lateral relaxation time scaling factor are both greater than 1 and are independent of each other.

[0096] In some embodiments, the system further includes a data weighting preprocessing module, which is used to divide the observed magnetic resonance signal into multiple time gates in chronological order before the start of the inversion iteration, count the noise amplitude in each time gate, take the reciprocal of the noise amplitude in each time gate to construct a data weighting matrix, and apply the data weighting matrix to the fitting residual vector in the observation response functional, so that the data of the high signal-to-noise ratio time gates obtain higher fitting priority during the inversion process.

[0097] In some embodiments, the joint regularization term in the objective function construction module is constructed using the minimum gradient support function, where each component of the minimum gradient support function is the sum of the square of the difference in physical properties between adjacent layers divided by the square and the square of the corresponding focusing scale factor; the joint regularization term is composed of a weighted sum of the water content focusing constraint term and the lateral relaxation time focusing constraint term.

[0098] In some embodiments, the iterative inversion solution module includes an iterative reweighting transformation submodule, which is used to construct a diagonal weight matrix in each iteration based on the gradient distribution of the inversion model vector obtained in the previous iteration step. The diagonal elements of the diagonal weight matrix are the reciprocals of the sum of the squares of the differences in physical properties between corresponding adjacent layers and the squares of the focusing scale factor. The iterative reweighting transformation submodule approximates the joint regularization term as a weighted quadratic form of the physical property gradient vector based on the diagonal weight matrix, so that the non-convex and non-quadratic inversion objective function is transformed into an analytically differentiable quadratic approximation form in the current iteration step.

[0099] In some embodiments, in the iterative inversion solution module, the momentum gradient descent method determines the update direction of the current iteration based on the gradient of the current approximate objective function and the momentum term of the previous iteration step in each iteration, and then corrects the inversion model vector along the update direction to form a new iterative model.

[0100] In some embodiments, the reservoir parameter extraction module includes a gradient extremum detection submodule. The gradient extremum detection submodule detects the location of local maxima on the water cut depth variation curve as a candidate depth for the top interface of the reservoir and detects the location of local minima as a candidate depth for the bottom interface of the reservoir. It also detects the location of local extrema on the lateral relaxation time depth variation curve as an independently verified interface candidate depth. When the difference between the interface locations determined by the two types of curves is less than a preset tolerance, the weighted average value is taken as the final interface burial depth.

[0101] In some embodiments, the reservoir parameter extraction module further includes a reservoir physical property parameter output submodule, which is used to read the water content value and lateral relaxation time value at the corresponding depth from the final inversion model vector according to the determined top and bottom interface burial depth, and calculate the statistical characteristics of the physical property parameters in the reservoir segment between the top and bottom interfaces, including the mean, median, maximum and minimum values.

[0102] The system runs on a high-performance computing platform and supports both single-point independent inversion mode and multi-point batch processing mode.

[0103] See Figure 2 As shown, the multi-scale focusing inversion results provided in the embodiments of this application, Figure 2 (a) in the figure is the water content inversion curve; Figure 2 (b) shows the inversion curve of lateral relaxation time. The red curve is the ideal curve of the ideal formation model, and the blue curve shows the inversion results as the formation depth changes. It can be seen that the overall trend of the blue curve is highly consistent with that of the red curve. Within the aquifer, the blue curve remains continuous and smooth, without obvious oscillations or fluctuations; at the top and bottom interfaces of the reservoir, the blue curve shows a steep gradient change, which is highly consistent with the position of the top and bottom plates of the aquifer corresponding to the red curve, and the interface response is clear and sharp. This indicates that the multi-scale focused inversion results have both intra-layer smoothness and inter-layer abrupt change characteristics, significantly enhancing the aquifer interface resolution capability. In terms of quantitative physical parameters, the water content and lateral relaxation time values ​​of the blue curve are close to the ideal values ​​of the red curve, with a small relative deviation, verifying the accuracy and reliability of the focused application in the quantitative evaluation of physical parameters.

[0104] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A magnetic resonance multi-scale focusing inversion method suitable for detecting salt lake brine, characterized in that, include: The observation magnetic resonance signal of the detection area is collected, the underground space of the detection area is divided into a one-dimensional layered medium model, the water content and transverse relaxation time of each discrete layer are used as the physical property parameters to be inverted, the inversion model vector is constructed, and a full waveform forward modeling operator is established to generate the predicted magnetic resonance signal corresponding to the current inversion model vector. An inversion objective function is constructed by superimposing an observation response functional and a joint regularization term. The observation response functional is used to measure the fitting residual between the predicted magnetic resonance signal and the observed magnetic resonance signal. The joint regularization term uses the difference in physical property parameters between adjacent layers in the inversion model vector as a gradient to apply a focusing constraint to the inversion model vector. The inversion objective function is transformed into an analytically differentiable approximation, and the momentum gradient descent method is used for iterative solution until the relative change in the model data fit difference satisfies the termination condition, and the final inversion model vector is output. Based on the final inversion model vector, the depth variation curves of water content and lateral relaxation time are extracted. The burial depth of the top and bottom interfaces of the salt lake brine reservoir and the corresponding physical parameters are determined according to the gradient extreme value positions of the depth variation curves.

2. The magnetic resonance multi-scale focusing inversion method for detecting salt lake brine according to claim 1, characterized in that, The inversion objective function is: ， In the formula, For the inversion objective function, To observe the response functional, For joint regularization terms, The inversion model vector is composed of water content. and lateral relaxation time Arrangement and composition, For full-waveform forward modeling analog operators, The difference in physical properties between adjacent strata, among which , , The number of strata. , For the first The water content of the first layer minus the first layer The moisture content of the layer, For the first The lateral relaxation time value of the layer minus the first Lateral relaxation time of the layer To observe magnetic resonance signals, Weighted matrix of data, These are the weighting coefficients of the joint normalization term. The weighting coefficients for the joint regularization term of water content are... The weight coefficients of the joint regularization term for the horizontal relaxation time; The focal scale factor for water content, The focal scale factor is the horizontal relaxation time.

3. The magnetic resonance multi-scale focusing inversion method for detecting salt lake brine according to claim 2, characterized in that, By dividing the original signal into multiple time gates and taking the reciprocal of the noise amplitude within each time gate, a data weighting matrix is ​​constructed.

4. The magnetic resonance multi-scale focusing inversion method for detecting salt lake brine according to claim 2, characterized in that, The focal scale factor of water content is in the first place. The value at the nth iteration is equal to the water content focusing scale factor scaling factor multiplied by the nth iteration. The median of the absolute values ​​of the water content differences between all adjacent layers in the inversion model vector obtained from the next iteration; The focusing scale factor of the transverse relaxation time in the first... The value at the nth iteration is equal to the scaling factor of the lateral relaxation time focus scale multiplied by the nth iteration. The median absolute value of the lateral relaxation time difference between all adjacent layers in the inversion model vector obtained from the next iteration.

5. The magnetic resonance multi-scale focusing inversion method for detecting salt lake brine according to claim 4, characterized in that, A transition factor associated with the number of inversion iterations is set, and the inversion process is coordinated to be in coarse-scale focus or fine-scale focus based on the transition factor. The weight coefficient of the joint regularization term is dynamically updated according to the transition factor.

6. The magnetic resonance multi-scale focusing inversion method for detecting salt lake brine according to claim 5, characterized in that, The transition factor associated with the number of inversion iterations is defined as follows: , For the first Iteration transition factor This is the preset number of focus transition iterations; The weight coefficients of the joint regularization term, dynamically updated according to the transition factor, are expressed as follows: ， In the formula, For the first The weight coefficients of the joint regularization term in the iteration. As the initial value, For the final value, when hour, The inversion is focused on a coarse scale. From initial value Increment to final value ;when hour, The inversion is focused on a fine scale. Fixed to final value .

7. The magnetic resonance multi-scale focusing inversion method for detecting salt lake brine according to claim 6, characterized in that, The fine-scale focusing constructs a gradient adaptive penalty based on the stratigraphic profile obtained from coarse-scale focusing. It dynamically adjusts the penalty intensity according to the ratio of differences in physical properties between adjacent stratigraphic units to the focusing scale factor, and sets a relative threshold coefficient. and , When the ratio of the difference in physical properties between adjacent strata to the focusing scale factor is less than When the ratio is greater than 1, configure the first penalty weight; when the ratio is greater than 1, configure the first penalty weight. When the ratio is greater than or equal to the first penalty weight, a second penalty weight is configured, and the second penalty weight is greater than the first penalty weight; and less than or equal to When the penalty weight is linearly interpolated between the first penalty weight and the second penalty weight, the first penalty weight and the second penalty weight are used as the pre-weighting coefficients in the joint regularization term.

8. The magnetic resonance multi-scale focusing inversion method for detecting salt lake brine according to claim 1, characterized in that, Transforming the inversion objective function into an analytically differentiable approximation includes: constructing a water content diagonal weight matrix and a lateral relaxation time diagonal weight matrix based on the gradient distribution of the inversion model vector, where the diagonal elements are the reciprocals of the sum of the squares of the differences in physical properties between corresponding adjacent layers and the squares of the focusing scale factor; and approximating the joint regularization term as a weighted quadratic form of the physical property gradient vector based on the water content diagonal weight matrix and the lateral relaxation time diagonal weight matrix, thereby transforming the original non-convex and non-quadratic inversion objective function into an analytically differentiable approximation.

9. A magnetic resonance multi-scale focusing inversion system suitable for detecting salt lake brine, characterized in that, include: The data acquisition and modeling module is used to acquire the observed magnetic resonance signals in the detection area, divide the underground space of the detection area into a one-dimensional layered medium model, construct the inversion model vector with the water content and lateral relaxation time of each discrete layer as the physical property parameters to be inverted, and establish a full waveform forward modeling operator to generate the predicted magnetic resonance signal corresponding to the current inversion model vector. The objective function construction module is used to construct an inversion objective function composed of the superposition of the observation response functional and the joint regularization term. The observation response functional is used to measure the fitting residual between the predicted magnetic resonance signal and the observed magnetic resonance signal. The joint regularization term uses the difference in physical property parameters between adjacent layers in the inversion model vector as a gradient to apply a focusing constraint to the inversion model vector. The iterative inversion solution module is used to transform the inversion objective function into an analytically differentiable approximate form, and to perform iterative solution using the momentum gradient descent method until the relative change in the model data fit difference satisfies the termination condition, and outputs the final inversion model vector. The reservoir parameter extraction module is used to extract the depth variation curves of water content and lateral relaxation time based on the final inversion model vector, and to determine the burial depth of the top and bottom interfaces of the salt lake brine reservoir and the corresponding physical property parameters based on the gradient extreme value positions of the depth variation curves.

Images