A continuous flow titration calorimeter incorporating microfluidics and machine learning correction and methods of use thereof

Abstract

The application provides a continuous flow dynamic titration calorimeter combining microfluidics and machine learning correction and a use method thereof. The device comprises: a heat preservation device for placing a calorimetry structure, preserving the calorimetry structure, collecting an ambient temperature of the calorimetry structure and sending the ambient temperature to a control and signal processing system; and the calorimetry structure comprises left and right material preheating aluminum blocks and a constant temperature aluminum block, a material access interface block A is arranged in a front groove of the material preheating aluminum block, a material access interface block B is arranged in a rear groove of the material preheating aluminum block, a thermoelectric power generation sheet A is arranged at a front part of the constant temperature aluminum block, a thermoelectric power generation sheet B is arranged at a rear part of the constant temperature aluminum block, and a thin film heating sheet is arranged above the thermoelectric power generation sheet A. The technical scheme of the application integrates microfluidic technology, thermoelectric power generation measurement and machine learning correction, and constructs a low-cost, miniaturized and high-precision continuous flow dynamic titration calorimetry method.

Description

Technical Field

[0001] This invention relates to the field of reaction heat measurement technology, and more particularly to a continuous flow dynamic titration calorimeter combining microfluidics and machine learning correction, and its method of use. Background Technology

[0002] Isothermal titration calorimetry (ITC), as a powerful analytical tool, is widely used in fields such as biochemistry, drug development, and materials science. It obtains thermodynamic parameters of a reaction, such as the binding constant, enthalpy change (ΔH), entropy change (ΔS), and stoichiometry, by measuring the heat released or absorbed when two substances are mixed at a constant temperature. Despite its crucial importance in experimental analysis, existing isothermal titration calorimeters face several limitations.

[0003] Traditional isothermal titration calorimeters are bulky, complex, and expensive, typically requiring sophisticated temperature control systems and highly sensitive calorimeters. For example, US Patent 9103782B2 describes a calorimeter with automatically adjustable titration rate, which improves the accuracy of the titration process, but its complex mechanical and electronic systems make the equipment expensive and difficult to maintain, unsuitable for resource-constrained environments. US Patent 8827549B2 demonstrates a multi-channel calorimeter, which increases experimental versatility, but its complex operation and large size make it unsuitable for scenarios requiring rapid and easy operation. While US Patent 10254239B2 focuses on using advanced algorithms to improve the speed and accuracy of data processing, it fails to address the issue of detection accuracy for low-concentration samples or weak thermal effects.

[0004] Furthermore, existing technologies are typically based on intermittent operation, meaning that after each titration, the system is allowed to reach thermal equilibrium before the next titration is performed, which has limited applicability in continuous processes. Regarding measurement accuracy, test results may contain significant errors due to factors such as ambient temperature fluctuations, baseline drift, and instrument noise. While US10337933B2 and US10429328B2 have improved measurement accuracy through improvements in detection technology and temperature control systems, their high cost and operational complexity remain. Meanwhile, the calorimeter with automated sample preparation and analysis described in US8449175B2 reduces the need for manual operation, but its high cost and reliance on technical personnel still limit its widespread application.

[0005] In recent years, the development of microfluidic technology has offered new possibilities for solving these problems. Microfluidic chips are widely used in various analytical instruments due to their small sample volume, high mass transfer efficiency, and ease of integration. For example, the microscale dynamic tracking reference continuous calorimetry device and method described in Chinese patent CN116818831A can perform rapid thermal effect detection in continuous flow conditions. However, this method is mainly suitable for processes with large amounts of heat release, and its ability to accurately capture small amounts of heat release is limited.

[0006] Given the limitations of existing isothermal titration calorimetry techniques in terms of cost, complexity, testing efficiency, and accuracy, there is an urgent need to develop a new calorimetric measurement method that is low-cost, high-precision, easy to operate, and suitable for continuous flow conditions. Summary of the Invention

[0007] To address the aforementioned technical problem that microfluidic technology cannot measure minute amounts of heat release in isothermal titration calorimetry, this invention provides a continuous flow dynamic titration calorimeter combining microfluidics and machine learning correction, along with its application method. This invention primarily constructs a low-cost, miniaturized, and high-precision continuous flow dynamic titration calorimetry method. This method can replace existing isothermal titration calorimetry methods, expanding the application scope of calorimetry.

[0008] The technical means employed in this invention are as follows:

[0009] A continuous flow dynamic titration calorimeter combining microfluidics and machine learning correction includes:

[0010] The heat preservation device is used to place the calorimetric structure, keep the calorimetric structure warm, collect the ambient temperature of the calorimetric structure, and send it to the control and signal processing system.

[0011] The calorimetric structure includes a material preheating aluminum block and a constant-temperature aluminum block arranged on the left and right sides. A material inlet / outlet interface block A is provided in the front groove of the material preheating aluminum block, and a material inlet / outlet interface block B is provided in the rear groove of the material preheating aluminum block. A thermoelectric generator A is provided at the front of the constant-temperature aluminum block, and a thermoelectric generator B is provided at the rear of the constant-temperature aluminum block. A thin-film heating element is provided above the thermoelectric generator A, a microfluidic chip A is provided above the thin-film heating element, and a microfluidic chip B is provided above the thermoelectric generator B.

[0012] The rear of the constant temperature aluminum block is connected to a constant temperature oil inlet, and the front of the constant temperature aluminum block is connected to a constant temperature oil outlet. The lower part of the material inlet / outlet interface block A is connected to material inlet A1, material inlet A2 and material outlet A respectively. The lower part of the material inlet / outlet interface block B is connected to material inlet B1, material inlet B2 and material outlet B respectively.

[0013] The first injection pump A, the second injection pump A, and the waste liquid tank are connected to the microfluidic chip A after passing through the material preheating aluminum block and the material inlet / outlet interface block A; the first injection pump B, the second injection pump B, and the waste liquid tank are connected to the microfluidic chip B after passing through the material preheating aluminum block and the material inlet / outlet interface block B.

[0014] The control and signal processing system is used to control the programmable DC power supply and the syringe pump controller; it is also used to collect signals generated by the thermoelectric generator and output calorimetric results.

[0015] Furthermore, the heat preservation device includes a calorimeter housing cover, a calorimeter housing base, and a thermocouple;

[0016] The calorimeter housing cover and calorimeter housing base are located outside the calorimeter structure, the thermal insulation felt cover is located outside the calorimeter housing cover, the thermal insulation felt base is located outside the calorimeter housing base, and the thermocouple is located inside the calorimeter housing cover.

[0017] Furthermore, silicone sealing rings are provided on the material inlet / outlet interface block A and the material inlet / outlet interface block B.

[0018] Furthermore, a material inlet / outlet viewing window is provided above the material inlet / outlet interface block A and the material inlet / outlet interface block B, and a microfluidic viewing window is provided above the microfluidic chip A and the microfluidic chip B.

[0019] Furthermore, the microfluidic chip A and the microfluidic chip B have the same structure, and the thermoelectric generator A and the thermoelectric generator B are of the same model.

[0020] Furthermore, the control and signal processing system includes a central control and data acquisition system, signal amplifier A, signal amplifier B, low-pass filter A, low-pass filter B, analog-to-digital converter, programmable DC power supply, and syringe pump controller;

[0021] The positive and negative electrodes of the programmable DC power supply are connected to the thin-film heating element via wires. The signal generated by the thermoelectric generator A is sequentially connected to the signal amplifier A, the low-pass filter A, and the analog-to-digital converter via a signal transmission line. The signal generated by the thermoelectric generator B is sequentially connected to the signal amplifier B, the low-pass filter B, and the analog-to-digital converter via a signal transmission line. The analog-to-digital converter transmits the signal to the central control and data acquisition system via the signal transmission line. The central control and data acquisition system controls the programmable DC power supply and the syringe pump controller via a communication line.

[0022] This invention also provides a method for using a continuous flow dynamic titration calorimeter that combines microfluidics and machine learning correction, based on any of the above-mentioned continuous flow dynamic titration calorimeters that combine microfluidics and machine learning correction, comprising the following steps:

[0023] S1. Materials are continuously added simultaneously to microfluidic chip A and microfluidic chip B at the same flow rate; wherein, in the test section, the material continuously introduced into the microchannel reference chip A through the first injection pump A is the titrant for the test titration process, and the material continuously introduced into the microchannel reference chip A through the second injection pump A is the titrant for the test titration process; in the reference section, the materials continuously introduced into the microfluidic chip B through the first injection pump B and the second injection pump B are both titrants; wherein the flow rates of the first injection pump A and the first injection pump B are the same, and the flow rates of the second injection pump A and the second injection pump B are the same.

[0024] S2. The target concentration of the solution in microfluidic chip A is controlled by controlling the flow rate of each injection pump; the titration process is simulated by adjusting the flow rate ratio of the injection pumps; constant temperature oil is introduced into the material preheating aluminum block and the constant temperature aluminum block through the constant temperature oil inlet, and the temperature of the constant temperature oil is the set test temperature; the signal values ​​generated by thermoelectric generator A and thermoelectric generator B are amplified by 100 times by signal amplifier A and signal amplifier B respectively, and then high frequency noise is filtered out by low-pass filter A and low-pass filter B respectively.

[0025] S3. Collect the signal values ​​of the two channels through an analog-to-digital converter and transmit them to the central control and data acquisition system;

[0026] S4. The central control and data acquisition system collects the temperature signals generated by thermocouples, cleans the signal data, and obtains a valid dataset.

[0027] S41. Under the same test conditions, continuously acquire the temperature signal generated by the thermocouple through the central control and data acquisition system for 10 minutes, and store the acquired temperature signal in the database, denoted as... Where T(k) represents the temperature signal value at the k-th sampling time, and N is the total number of sampling points;

[0028] S42. Perform a Fast Fourier Transform on the acquired temperature signal data to obtain the energy distribution of the signal in the frequency domain. The time-domain signal {T(k)} can be transformed to the frequency domain as follows:

[0029]

[0030] Where ω is the frequency sampling point index (0≤ω) <N);

[0031] Remove DC component: Set the component (DC quantity) corresponding to ω = 0 to 0; Find the maximum peak value: Within the range of ω > 0, find the frequency component ω with the largest amplitude. max The corresponding frequency is obtained by conversion. Where f sIndicates the signal sampling frequency; set the cutoff frequency: 1.5 × f max The cutoff frequency of the Butterworth low-pass filter is denoted as f. cut =1.5·f max ;

[0032] S43. A Butterworth low-pass filter is used to filter the temperature signal {T(k)} to remove high-frequency noise. The transfer function of the filter can be expressed as:

[0033]

[0034] Where, ω c =2πf cut Let {T} be the angular cutoff frequency of the filter. By discretizing the transfer function, it is mapped to the z-domain, thus realizing digital filtering. The filtered data is denoted as {T}. flt1 (k)}, thus obtaining the preliminary filtering results;

[0035] S44, Based on the preliminary filtering results {T flt1 (k)}, further employing a two-dimensional Kalman filter to estimate and correct the measurement signal; the system state model is:

[0036] x k =Ax k-1 +Bu k +w k

[0037] Where, x k Let u represent the state vector at time k. k w is an optional control vector. k Let A represent process noise, and A and B be the system state transition matrix and control matrix, respectively; the observation model is:

[0038] z k =Hx k +v k

[0039] Among them, z k For the observed value, v k For observation noise, H is the observation matrix; in the Kalman filter recursive equation, the prediction stage is:

[0040]

[0041] P k|k-1 =AP k|k-1 A T +Q

[0042] Among them, P k|k-1 For state covariance prediction, Q is the process noise covariance matrix, and the update stage is:

[0043] K k =P k|k-1 H T HP k|k-1 H T +R) -1

[0044]

[0045] P k|k =(IK k H)P k|k-1

[0046] Among them, K k Let R be the Kalman gain matrix and R be the observation noise covariance matrix. The final temperature signal {T} is obtained after two-dimensional Kalman filtering correction. flt2 (k)};

[0047] S45. Record the heat dissipation power data of the thin-film heating element as follows: The data is then unfolded into a one-dimensional data sequence in chronological order to maintain the comparability of heat release power at different time points; P(k) and the sampling time T of the temperature signal are also considered. flt2 (k) One-to-one correspondence;

[0048] S46. Use a window function to perform piecewise analysis on {P(k)}; set the window width W = 60 and the window movement step size s = 2; statistically analyze the power data in the m-th window, where the data in the window is {P(k)|k∈[k]}. start ,k end ]} Calculate the first quartile Q1 and the third quartile Q3 within the window, where Q1 represents the value at the 25th percentile after sorting the data in ascending order, and Q3 represents the value at the 75th percentile after sorting the data in ascending order;

[0049] S47. For each window's data {P(k)}, perform the following checks to calculate the interquartile range (IQR) within that window:

[0050] IQR = Q3 - Q1

[0051] To determine if the data is an outlier, if P(k) exceeds the interval [Q1-1.5×IQR, Q3+1.5×IQR], it is considered an outlier; all heat release power values ​​marked as outliers and their corresponding measurement signal data {T} are then analyzed. flt2 (k)} are all removed, retaining only the normal data and its temperature signal pairs; through the above window function traversal and anomaly marking, extreme noise points that occur during observation or acquisition are removed, thereby obtaining the effective dataset {(T flt2 (k),P(k))};

[0052] S5. Employ machine learning-based methods to build a meta-model based on an effective dataset;

[0053] S6. Use a meta-model to predict the heat of reaction.

[0054] Furthermore, S5 specifically includes the following steps:

[0055] S51. Three algorithms—XGBoost, LightGBM, and Random Forest—are used to model and train the training data, respectively. Parameters are tuned using methods such as grid search, random search, or Bayesian optimization to obtain relatively stable parameter configurations, thus forming three sets of base models. The objective function of XGBoost is:

[0056]

[0057] Among them, y i Indicates the true label, For the cumulative prediction of the first m-1 trees, f m Let Ω(·) represent the residual function learned by the m-th tree, and let Ω(·) be the tree complexity regularization term.

[0058] For the m-th tree, the output of the LightGBM model can be approximated as:

[0059] F m (x)=F m-1 (x)+α m T m (x)

[0060] Where, α m Let T be the learning rate of the m-th tree. m (x) represents the estimate of the input x by the m-th tree; the random forest obtains the final prediction output by constructing K CART decision trees and performing voting or averaging. The model can be represented as:

[0061]

[0062] Among them, T k (x) represents the output of the k-th decision tree;

[0063] S52. After tuning the hyperparameters of XGBoost, LightGBM, and Random Forest respectively, the optimal combination of hyperparameters was finally locked and used to train three base models, denoted as:

[0064] M XGB M LGB M RF

[0065] S53. Input the same training or validation set data into the three base models in sequence to obtain three sets of predicted values:

[0066]

[0067] Where X represents the input feature matrix, Let X represent the predicted heat release vectors output by XGBoost, LightGBM, and Random Forest, respectively. By comparing the predicted heat release vectors with the corresponding true values ​​Y, we can obtain the error and residual information generated by each of the three models.

[0068] S54. Use the true value Y as the training label for the meta-model, and... The new feature vector is formed by concatenating or combining the features, denoted as:

[0069]

[0070] S55. Support vector machine regression is used as the meta-model, denoted as M. SVM Its objective function can take the form of ∈-insensitive loss:

[0071]

[0072] Where w and b are the weight vector and bias term of the support vector machine, respectively, C is the regularization coefficient, and ξ is the bias term. i Represents slack variables that extend beyond the -insensitive interval; kernel function K(x) i ,x j The kernel can be selected from RBF kernel, linear kernel or polynomial kernel form according to actual needs;

[0073] S56. Utilize cross-validation, hold-out validation, or other validation techniques to adjust and optimize the various parameters of the SVM, including the kernel function type, regularization coefficient C, and kernel function parameter γ; specifically, it can be expressed as:

[0074]

[0075] Here, CVBloss represents the comprehensive loss value obtained by measuring and averaging the prediction error during cross-validation; through this process, the meta-model fully learns the mapping relationship between the base model output and the true label;

[0076] S57. When the prediction error of the meta-model on the validation set meets the preset accuracy requirements, the parameter determination of the meta-model is completed, and the final meta-model is obtained.

[0077] Furthermore, when the ambient temperature during actual measurement exceeds the temperature range of the machine learning training set, a method is used to improve the robustness of the model output in extrapolation scenarios through linear mapping to map the temperature outside the temperature range to the temperature range. The steps are as follows:

[0078] P1. Based on the correspondence between temperature T and measurement signal M in the existing training dataset, the linear fitting model is obtained using the least squares method or other linear regression methods, denoted as:

[0079] M = k·T + b

[0080] Where M represents the measurement signal, and k and b are the slope and intercept of the linear model, respectively, as shown in the following formula:

[0081]

[0082] Among them, T i and M i To obtain the temperature and measurement signal of the i-th sample in the training set, and These are the average values ​​of the temperature and the measured signal, respectively.

[0083] P2. Set the temperature range for the training data to [T]. min ,T max This interval can be considered as the core region where the model has high predictive reliability in the temperature dimension; {k,b} and T min ,T max Stored in the parameter library for later use;

[0084] P3. For the new test dataset in Let j be the temperature value of the j-th test sample. Other feature vectors or measurement signals representing this sample need to be determined first. Does it fall under [T]? min ,T max Within the range; if If so, the temperature is considered normal and no correction is needed; or These are then marked as samples with temperatures outside the range and added to set O;

[0085] P4. For samples marked as out of range The linear relationship is invoked to correct the corresponding measurement signal; assuming the original value of the measurement signal for out-of-range samples is... The mapping correction formula is:

[0086]

[0087] in, This indicates the corrected measurement signal, in conjunction with... Ensure new test samples Closer to the training range [T] min ,T maxThe corresponding feature distribution;

[0088] P5. Replace all the original out-of-range data with the linearly corrected test samples to form a new test dataset. At this point, the existing meta-model can be used to make predictions on the updated test data.

[0089] Furthermore, the specific steps of S6 are as follows:

[0090] S61. Determine whether the process is endothermic or exothermic. If the process is exothermic, execute S62 to S65. If the process is exothermic, execute S66, S67, S63, S64, and S65 in sequence.

[0091] S62. Input the preprocessed measurement signal, which has been collected in the early stage and filtered and removed outliers, into the machine learning model. The meta-model outputs the predicted value of the heat release power of the process, and denots this predicted power as P. 预测 =P;

[0092] S63. Substitute the predicted heat release power P output by the machine learning model into the following formula to calculate the heat release Q of the process:

[0093]

[0094] Where V is the liquid volume in the microfluidic chip, representing the volume of liquid stored in the chip during the titration process or other testing processes; q is the flow rate of the material in the microfluidic channel, which is usually set and monitored by a microfluidic pump or other precision flow rate control device; and c is the concentration of the target substance in the test material, used for the subsequent calculation of the heat of reaction ΔH.

[0095] S64. Given that the target concentration of the tested material is c, the total feed amount or reactant content can be characterized under the condition that the microfluidic chip volume is V; convert μL to L.

[0096] S65. Based on the measured heat of release Q and concentration c, the heat of reaction ΔH is calculated:

[0097]

[0098] S66. If the process is determined to be an endothermic process, then an additional exothermic power P needs to be provided to the thin-film heating element through a programmable DC power supply. 补偿 And ensure that the compensation value can cover the power absorbed during the heat absorption process, so as to ensure that the whole system is in the desired temperature or energy balance state;

[0099] S67, based on the predicted heat absorption power P and the compensation power P 补偿 Determine the actual required heating element power:

[0100] P = P 补偿 -P 预测

[0101] Among them, P 预测 P represents the endothermic power predicted by the machine learning model. 补偿 It needs to be greater than |P 预测 |, to meet the energy compensation requirements of the external heat absorption process.

[0102] Compared with the prior art, the present invention has the following advantages:

[0103] 1. Reduced Costs and Simplified Structure: This invention uses a microfluidic chip, a thermoelectric generator, and a signal amplifier as core components, significantly reducing the instrument's manufacturing cost and size. The detachable design of the microfluidic chip allows for flexible adjustment of the hybrid structure according to actual needs, facilitating later maintenance and replacement. This makes the calorimeter more suitable for use in resource-constrained laboratory or field environments, broadening its application scope.

[0104] 2. Improved measurement accuracy and reduced errors: Addressing the complex error factors affecting accuracy during testing (such as flow rate, test temperature, ambient temperature, and signal values ​​at the reference and test ends), this invention introduces machine learning methods to establish a precise calibration model. Through precise calibration of the calorimeter, test errors are significantly reduced, and detection accuracy is improved, particularly in the detection of low-concentration samples and weak thermal effects, solving the problem of inaccurate measurement using traditional methods.

[0105] 3. Achieving dynamic titration in continuous flow: A microfluidic chip is used as the carrier for continuous material flow and mixing, combined with a high-precision syringe pump to precisely adjust the flow rate ratio, dynamically simulating the titration process in continuous flow. This invention can obtain the thermodynamic parameters of substances in a continuous state, overcoming the shortcomings of traditional batch methods and meeting the needs of specific application scenarios such as continuous process control and real-time monitoring.

[0106] 4. Simple operation, high efficiency and speed: This invention simplifies the detection process through direct energy conversion and measurement, reducing reliance on complex mechanical and electronic systems. It eliminates the need for sophisticated temperature control systems and highly sensitive traditional heat detectors, lowering the requirements for professional operators, improving testing efficiency, and making it suitable for scenarios requiring rapid and easy operation.

[0107] 5. Expanding the Application Scope of Calorimetry: This invention transforms extremely weak thermal signals into usable information with a high signal-to-noise ratio through the cascaded processing of a signal amplifier and a low-pass filter. On the one hand, the signal amplifier significantly amplifies the extremely small thermal signal, avoiding the loss of effective information due to insufficient quantization accuracy during subsequent digitization. On the other hand, the low-pass filter removes high-frequency noise or jitter from the amplified signal, making the signal input to the analog-to-digital converter purer and more stable, thereby greatly improving the accuracy and resolution of micro-thermal change detection. The innovative application of machine learning technology and linear mapping correction methods maintains high prediction accuracy and stability even when the model is extrapolated and the temperature deviates from the training range. The overall device combines a microfluidic chip with a thermoelectric generator, offering advantages such as compact structure, miniaturization, and low energy consumption. Furthermore, combined with intelligent machine learning algorithms, it can flexibly adapt to various chemical systems or reaction conditions, demonstrating broad applicability and scalability for various titration calorimetric needs.

[0108] 6. The present invention provides a heat preservation device for housing and heat preservation of a calorimetric structure; the heat preservation device can also collect the ambient temperature of the calorimetric structure and send the collected temperature signal to a control and signal processing system in real time. By placing the calorimetric structure inside the heat preservation device, the interference caused by external ambient temperature fluctuations on the main body of the calorimetric device can be effectively reduced, thereby significantly improving the accuracy and stability of the measurement of microcalorimetric changes.

[0109] In summary, this invention constructs an innovative continuous-flow dynamic titration calorimetry method by combining microfluidic technology, thermoelectric power generation measurement, and machine learning correction. This method not only effectively replaces existing isothermal titration calorimetry techniques but also achieves comprehensive improvements in cost, accuracy, efficiency, and applicability, demonstrating significant practical application value and promising prospects for widespread adoption. Attached Figure Description

[0110] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0111] Figure 1 This is a schematic diagram of the calorimeter structure of the present invention.

[0112] Figure 2 This is an exploded view of the calorimeter of the present invention.

[0113] Figure 3 This is a calorimetric structural diagram of the present invention.

[0114] Figure 4 This is a diagram of the control and signal processing system of the present invention.

[0115] Figure 5 This is a flowchart of the signal value processing of the present invention.

[0116] Figure 6 This is the machine learning-assisted calorimetric correction diagram of the present invention.

[0117] Figure 7 This is a diagram showing the error distribution of the calorimeter used in this invention.

[0118] Figure 8 This is a diagram illustrating the fitting process in an embodiment of the present invention.

[0119] Figure 9 This is a data plot for Embodiment 1 of the present invention.

[0120] Figure 10 This is a data plot for Embodiment 2 of the present invention.

[0121] Figure 11 This is a data plot for Embodiment 3 of the present invention.

[0122] Figure 12 This is a data plot for Embodiment 4 of the present invention.

[0123] In the diagram: 11. Insulating felt top cover; 12. Insulating felt base; 21. Calorimeter outer shell top cover; 22. Calorimeter outer shell base; 3. Thermocouple; 411. Material inlet / outlet viewing window; 412. Microfluidic viewing window; 421. Microfluidic chip A; 422. Microfluidic chip B; 43. Thin-film heating element; 441. Thermoelectric generator A; 442. Thermoelectric generator B; 451. Material inlet / outlet interface block A; 452. Material inlet / outlet interface block B; 453. Silicone sealing ring; 461. Material preheating aluminum block; 462. Constant temperature aluminum block; 463. Constant temperature oil inlet; 464. Constant temperature oil outlet; 471. Material Inlet A1; 472. Material Inlet A2; 473. Material Outlet A; 474. Material Inlet B1; 475. Material Inlet B2; 476. Material Outlet B; 481. First Injection Pump A; 482. Second Injection Pump A; 483. First Injection Pump B; 484. Second Injection Pump B; 49. Waste Liquid Tank; 51. Central Control and Data Acquisition System; 521. Signal Amplifier A; 522. Signal Amplifier B; 531. Low-Pass Filter A; 532. Low-Pass Filter B; 54. Analog-to-Digital Converter; 55. Programmable DC Power Supply; 56. Injection Pump Controller. Detailed Implementation

[0124] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0125] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0126] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0127] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0128] In the description of this invention, it should be understood that the orientation or positional relationship indicated by directional terms such as "front, back, up, down, left, right", "horizontal, vertical, horizontal" and "top, bottom" is generally based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing this invention and simplifying the description. Unless otherwise stated, these directional terms do not indicate or imply that the device or element referred to must have a specific orientation or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on the scope of protection of this invention. The directional terms "inner" and "outer" refer to the inner and outer contours relative to the outline of each component itself.

[0129] For ease of description, spatial relative terms such as "above," "over," "on the upper surface of," "above," etc., are used herein to describe the spatial positional relationship of a device or feature as shown in the figures to other devices or features. It should be understood that spatial relative terms are intended to encompass different orientations in use or operation besides the orientation of the device as described in the figures. For example, if the device in the figures is inverted, a device described as "above" or "above" other devices or structures would subsequently be positioned as "below" or "under" other devices or structures. Thus, the exemplary term "above" can include both "above" and "below." The device may also be positioned in other different ways (rotated 90 degrees or in other orientations), and the spatial relative descriptions used herein will be interpreted accordingly.

[0130] Furthermore, it should be noted that the use of terms such as "first" and "second" to define components is merely for the purpose of distinguishing the corresponding components. Unless otherwise stated, the above terms have no special meaning and therefore should not be construed as limiting the scope of protection of this invention.

[0131] like Figure 1-4 As shown, this invention provides a continuous flow dynamic titration calorimeter combining microfluidics and machine learning correction, comprising:

[0132] A heat preservation device is used to place the calorimeter structure, keep the calorimeter structure warm, collect the ambient temperature of the calorimeter structure and send it to the control and signal processing system; the heat preservation device includes a calorimeter outer shell cover 21, a calorimeter outer shell base 22 and a thermocouple 3.

[0133] The calorimeter outer cover 21 and the calorimeter outer base 22 are located outside the calorimeter structure, the thermal insulation felt cover 11 is located outside the calorimeter outer cover 21, the thermal insulation felt base 12 is located outside the calorimeter outer base 22, and the thermocouple 3 is located inside the calorimeter outer cover 21.

[0134] The calorimetric structure includes a material preheating aluminum block 461 and a constant-temperature aluminum block 462 arranged on the left and right sides. A material inlet / outlet interface block A451 is located in the front groove of the material preheating aluminum block 461, and a material inlet / outlet interface block B452 is located in the rear groove of the material preheating aluminum block 461. A thermoelectric generator A441 is located at the front of the constant-temperature aluminum block 462, and a thermoelectric generator B442 is located at the rear of the constant-temperature aluminum block 462. A thin-film heating element 43 is located above the thermoelectric generator A441, a microfluidic chip A421 is located above the thin-film heating element 43, and a microfluidic chip B422 is located above the thermoelectric generator B442. The left end is a reference section, and the right end is a testing section. Silicone sealing rings 453 are provided on the material inlet / outlet interface blocks A451 and B452. A material inlet / outlet viewing window 411 is provided above the material inlet / outlet interface block A451 and the material inlet / outlet interface block B452, and a microfluidic viewing window 412 is provided above the microfluidic chip A421 and the microfluidic chip B422. The microfluidic chip A421 and the microfluidic chip B422 have the same structure, and the thermoelectric generator A441 and the thermoelectric generator B442 are of the same model.

[0135] The rear of the constant temperature aluminum block 462 is connected to a constant temperature oil inlet 463, and the front of the constant temperature aluminum block 462 is connected to a constant temperature oil outlet 464. The lower part of the material inlet / outlet interface block A451 is connected to the material inlet A1471, the material inlet A2472 and the material outlet A473 respectively. The lower part of the material inlet / outlet interface block B452 is connected to the material inlet B1474, the material inlet B2475 and the material outlet B476 respectively.

[0136] The first injection pump A481, the second injection pump A482, and the waste liquid tank 49 are connected to the microfluidic chip A421 after passing through the material preheating aluminum block 461 and the material inlet / outlet interface block A451; the first injection pump B483, the second injection pump B484, and the waste liquid tank 49 are connected to the microfluidic chip B422 after passing through the material preheating aluminum block 461 and the material inlet / outlet interface block B452.

[0137] A control and signal processing system is used to control the programmable DC power supply 55 and the syringe pump controller 56; and to collect signals generated by the thermoelectric generator and output calorimetric results.

[0138] The control and signal processing system includes a central control and data acquisition system 51, signal amplifier A521, signal amplifier B522, low-pass filter A531, low-pass filter B532, analog-to-digital converter 54, programmable DC power supply 55, and injection pump controller 56.

[0139] The positive and negative electrodes of the programmable DC power supply 55 are connected to the thin-film heating element 43 via wires. The signal generated by the thermoelectric generator A441 is connected to the signal amplifier A521, the low-pass filter A531, and the analog-to-digital converter 54 via a signal transmission line. The signal generated by the thermoelectric generator B442 is connected to the signal amplifier B522, the low-pass filter B532, and the analog-to-digital converter 54 via a signal transmission line. The analog-to-digital converter 54 transmits the signal to the central control and data acquisition system 51 via a signal transmission line. The central control and data acquisition system 51 controls the programmable DC power supply 55 and the syringe pump controller 56 via a communication line.

[0140] This invention also provides a method for using a continuous flow dynamic titration calorimeter that combines microfluidics and machine learning correction, comprising:

[0141] S1. Microfluidic chip A and microfluidic chip B have identical structures. Materials are continuously and simultaneously added to microfluidic chip A and microfluidic chip B at the same flow rate. In the test section, the materials continuously introduced into the microchannel reference chip A via the first injection pump A and the second injection pump A are the titrant and the titrant being tested, respectively. In the reference section, the materials continuously introduced into the microfluidic chip B via the first injection pump B and the second injection pump B are both the titrant being tested. The flow rates of the first injection pump A and the first injection pump B are the same, and the flow rates of the second injection pump A and the second injection pump B are also the same.

[0142] S2. The target concentration of the solution in microfluidic chip A is controlled by adjusting the flow rate of each syringe pump. The titration process is simulated by adjusting the flow rate ratio of the syringe pumps. Constant temperature oil is introduced into the material preheating aluminum block and the constant temperature aluminum block through the constant temperature oil inlet 463; the temperature of the constant temperature oil is the set test temperature. Figure 5 The signal values ​​generated by thermoelectric generator A and thermoelectric generator B are amplified by 100 times by signal amplifier A and signal amplifier B, respectively, and then high-frequency noise is filtered out by low-pass filter A and low-pass filter B, respectively.

[0143] S3. Then, the signal values ​​of the two channels are collected through an analog-to-digital converter and transmitted to the central control and data acquisition system.

[0144] S4. Simultaneously, the central control and data acquisition system collects the temperature signals generated by the thermocouples; the signal data is cleaned to obtain a valid dataset.

[0145] S41. Under the same test conditions, continuously acquire the temperature signal generated by the thermocouple through the central control and data acquisition system for 10 minutes, and store the acquired temperature signal in the database, denoted as... Among them, T(k) represents the temperature signal value at the k-th sampling moment, and N is the total number of sampling points. To ensure the consistency of the test conditions, this step can be carried out multiple times under the same ambient temperature and the same operation process, so as to obtain several groups of temperature signal data for subsequent analysis.

[0146] S42. Perform a fast Fourier transform (FFT) on the collected temperature signal data to obtain the energy distribution of the signal in the frequency domain. Specifically, the transformation of the time-domain signal {T(k)} to the frequency domain can be expressed as:

[0147]

[0148] where ω is the frequency sampling point index (0 ≤ ω < N). Remove the DC component: that is, set the component corresponding to ω = 0 (DC quantity) to 0; find the maximum peak: within the range of ω > 0, find the frequency component ω with the largest amplitude max , and convert it to the corresponding frequency where f s represents the signal sampling frequency; set the cut-off frequency: take 1.5×f max as the cut-off frequency of the Butterworth low-pass filter, denoted as f cut = 1.5·f max .

[0149] S43. Use a Butterworth low-pass filter to filter the temperature signal {T(k)} to remove high-frequency noise. The transfer function of the filter can be expressed as:

[0150]

[0151] where ω c = 2πf cut is the angular cut-off frequency of the filter. In this invention, the cut-off frequency is 0.5. Through discretization (such as bilinear transformation), this transfer function can be mapped to the z-domain to achieve digital filtering. The filtered data is denoted as {T flt1 (k)}, and the preliminary filtering result is obtained.

[0152] S44. Based on the preliminary filtering result {T flt1 (k)}, further use two-dimensional Kalman filtering to estimate and correct the measurement signal. Two-dimensional Kalman filtering jointly estimates the signal in the time dimension and an additional spatial or parameter dimension, which can effectively reduce the influence of system noise and observation noise, so as to obtain a measurement signal closer to the true value. The system state model is:

[0153] x k = Ax k-1 + Bu k + w k

[0154] Where, x k u represents the state vector at time k (which may include temperature and additional states). k w is an optional control vector. k Let be the process noise, and a and B be the system state transition matrix and control matrix, respectively. The observation model is:

[0155] z k =Hx k +v k

[0156] Among them, z k The observed value (here referring to the filtered temperature measurement value {T}) flt1 (k)} and corresponding spatial / parameter information), v k Let H be the observation noise and H be the observation matrix. In the Kalman filter recursive equation, the prediction stage is:

[0157]

[0158] P k|k-1 =AP k|k-1 A T +Q

[0159] Among them, P k|k-1 For state covariance prediction, Q is the process noise covariance matrix. The update phase is as follows:

[0160] K k =P k|k-1 H T HP k|k-1 H T +R) -1

[0161]

[0162] P k|k =(iK k H)P k|k-1

[0163] Among them, K k Let R be the Kalman gain matrix and R be the observation noise covariance matrix. The final temperature signal {T} after two-dimensional Kalman filtering correction is obtained. flt2 (k)}.

[0164] S45. Record the heat dissipation power data of the thin-film heating element as follows: The data is then unfolded into a one-dimensional data sequence in chronological order to maintain the comparability of heat release power at different time points. Here, P(k) is related to the sampling time T of the temperature signal. flt2 (k) One-to-one correspondence.

[0165] S46. Use a window function to perform piecewise analysis on {P(k)}. Set the window width W = 60 and the window movement step size s = 2. Statistically analyze the power data within the m-th window, where the data within the window is {P(k)|k∈[k...]}. start ,k end ]} Calculate the first quartile Q1 and the third quartile Q3 within the window, where Q1 represents the value at the 25th percentile after sorting the data in ascending order, and Q3 represents the value at the 75th percentile after sorting the data in ascending order.

[0166] S47. For each window's data {P(k)}, perform the following judgment: Calculate the interquartile range (IQR) within that window:

[0167] IQR = Q3 - Q1

[0168] To determine if the data is an outlier, if P(k) exceeds the interval [Q1-1.5×IQR, Q3+1.5×IQR], it is considered an outlier. All heat release power values ​​marked as outliers and their corresponding measurement signal data {T} are then analyzed. flt2 (k)} are all removed, retaining only the normal data and its temperature signal pairs. Through the above window function traversal and anomaly labeling, extreme noise occurring during observation or acquisition can be effectively removed, thus obtaining a new, effective dataset {(T}. flt2 This dataset contains signal values ​​from both the reference and test components, as well as information such as the ambient temperature of the device, providing a more accurate and clean foundation for subsequent feature extraction and modeling.

[0169] Each test involved continuously acquiring signal values ​​for 10 minutes under identical conditions. The collected data was further processed by a central control and data acquisition system. First, outliers were removed using the interquartile range (IQR) method, specifically the data from the first 5 minutes. The stable signal values ​​from the remaining 5 minutes were then subjected to Kalman filtering to further eliminate noise interference. Another round of IQR was used to remove outliers, and the average of the last 2 minutes of data was calculated. This yielded the signal values ​​for the reference and test portions under the specified test conditions, as well as the ambient temperature of the device.

[0170] S5. Employ machine learning-based methods to build a meta-model based on an effective dataset;

[0171] S51. After data preprocessing, three algorithms—XGBoost, LightGBM, and RandomForest—are used to model and train the training data. Parameter tuning (hereinafter referred to as "parameter tuning") is performed using methods such as grid search, random search, or Bayesian optimization to obtain relatively stable parameter configurations, thus forming three sets of base models. XGBoost is an efficient implementation based on the Gradient Boosting Decision Tree (GBDT) framework, and its objective function can be expressed as:

[0172]

[0173] Among them, y i Indicates the true label, For the cumulative prediction of the first m-1 trees, f m Let Ω(·) represent the residual function learned by the m-th tree, and let Ω(·) be the tree complexity regularization term. By iteratively training multiple decision trees, XGBoost can effectively capture the non-linear structure in the data. LightGBM is also based on the GBDT framework, but it adopts an efficient data structure based on gradient histograms and an incremental learning strategy based on leaf-wise approaches. For the m-th tree, the output of the LightGBM model can be approximately represented as:

[0174] F m (x)=F m-1 (x)+α m T m (x)

[0175] Where, α m Let T be the learning rate of the m-th tree. m (x) represents the estimate of the input x by the m-th tree. Random forests obtain the final prediction output by constructing K CART decision trees and performing voting or averaging (generally the mean is taken in regression scenarios). The model can be represented as:

[0176]

[0177] Among them, T k (x) represents the output of the k-th decision tree. Since each decision tree is independent and the training data is randomly sampled and features are randomly selected, random forests have strong generalization ability and resistance to overfitting.

[0178] S52. After tuning the hyperparameters of XGBoost, LightGBM, and Random Forest respectively, the optimal combination of hyperparameters was finally locked and used to train three base models, denoted as:

[0179] M XGB M LGB M RF

[0180] S53. Input the same training set (or validation set) data into the three base models in sequence to obtain three sets of predicted values:

[0181]

[0182] Where X represents the input feature matrix. Let Y represent the predicted heat release vectors output by XGBoost, LightGBM, and Random Forest, respectively. By comparing these predicted values ​​with the corresponding true values ​​Y, we can obtain the error and residual information generated by each of the three models, providing a basis for subsequent meta-model training.

[0183] To fully explore the differences and complementarities in how different base models learn data features, this invention constructs a meta-model to further learn and fuse the three sets of prediction results. This meta-model can, to a certain extent, capture the biases, residual distributions, and learning preferences of the three base models for different input features, thereby further improving the overall prediction accuracy and stability.

[0184] S54. Use the true value Y as the training label (true value) of the meta-model, and... The new feature vector is formed by concatenating or combining the features, denoted as:

[0185]

[0186] S55. Support Vector Machine Regression (SVM Regression) is used as the meta-model, denoted as M. SVM Its objective function can take the form of ∈-insensitive loss:

[0187]

[0188] In this formula, w and b are the weight vector and bias term of the support vector machine, respectively, C is the regularization coefficient, and ξ is the bias term. i This represents a slack variable that extends beyond the -insensitive interval. The kernel function K(x) i ,x j The kernel can be selected from RBF kernel, linear kernel or polynomial kernel, etc., according to actual needs.

[0189] S56. Using cross-validation, hold-out validation, or other validation techniques, the parameters of the SVM (including kernel type, regularization coefficient C, kernel parameter γ, etc.) are adjusted and optimized. Specifically, this can be expressed as:

[0190]

[0191] Here, CVBloss represents the comprehensive loss value obtained by measuring and averaging the prediction error (such as mean squared error, MSE) during cross-validation (e.g., K-fold cross-validation). This process enables the meta-model to fully learn the mapping relationship between the base model output and the true label.

[0192] S57. When the prediction error of the meta-model on the validation set meets the preset accuracy requirement (e.g., MSE≤δ), the parameter determination of the meta-model is completed, and the final meta-model is obtained. This meta-model approximates the outputs of three base models in a high-dimensional feature space, which can effectively improve the overall prediction accuracy and stability.

[0193] By using a meta-model to learn the outputs of the three base models, the differences and complementarities between the different models in terms of local features and residual information are fully captured, thereby improving the overall prediction accuracy. The three base models learn for different feature distributions or training preferences, and the meta-model is then fused on this basis, thus significantly enhancing the system's adaptability to diverse scenarios.

[0194] This invention addresses the problem of limited temperature range in training data by proposing a method to improve the robustness of model output in extrapolation scenarios through linear mapping. In traditional modeling, when the temperature value of the test data exceeds the training temperature range, the model is prone to overfitting or misfitting, leading to significant deviations in prediction results. To address this, this invention constructs a linear fitting model of temperature and measurement signal based on existing training data beforehand, and uses this linear relationship as a "mapping function" to correct test samples with temperatures exceeding the range, thereby effectively improving the stability and accuracy of the model during extrapolation.

[0195] P1. Based on the correspondence between temperature T and measured signal M in the existing training dataset, the linear fitting model is obtained using the least squares method or other linear regression methods, denoted as P1.

[0196] M = k·T + b

[0197] Where M represents the measurement signal (or other relevant parameters may be used), and k and b are the slope and intercept of the linear model, respectively, given by the following equation:

[0198]

[0199] Among them, T i and M i To obtain the temperature and measurement signal of the i-th sample in the training set, and These are the average values ​​of temperature and the measured signal, respectively.

[0200] P2. Set the temperature range for the training data to [T]. min ,Tmax This interval can be considered as the core region where the model has high predictive reliability in the temperature dimension. Let {k,b} and T... min ,T max Stored in the parameter library for later use.

[0201] P3. For the new test dataset in Let j be the temperature value of the j-th test sample. Other feature vectors or measurement signals representing this sample need to be determined first. Does it fall under [T]? min ,T max Within the range. Specifically, if If so, the temperature is considered normal and no correction is needed; or Then it is marked as "temperature out of range sample" and recorded in set O.

[0202] P4. For samples marked as out of range The linear relationship recorded above is used to correct the corresponding measurement signal. Let the original value of the measurement signal for the out-of-range sample be... (If the measurement signal is not yet clear, it can be obtained from an existing model or sensor readings), then the mapping correction formula is:

[0203]

[0204] in, This indicates the corrected measurement signal. (In conjunction with...) This ensures the availability of new test samples. Closer to the training range [T] min ,T max The characteristic distribution corresponding to ].

[0205] P5. Replace all the original out-of-range data with the linearly corrected test samples to form a new test dataset. At this point, the existing prediction model can be used to predict the updated test data, thereby maintaining high prediction accuracy and stability in temperature extrapolation scenarios.

[0206] When the test data temperature falls outside the training range, the measured signal is corrected by linear fitting, which can significantly reduce prediction fluctuations and mismatches in the temperature extrapolation region, thereby improving the stability and accuracy of the prediction. The linear mapping required by this invention only requires a simple correction of {k,b} lines, which can significantly reduce costs in terms of algorithm complexity and time overhead. This linear mapping strategy can be combined with various models (such as tree-based methods, neural networks, linear regression, etc.) and can also be extended to other extrapolation variable scenarios (such as pressure, humidity, etc.), showing strong versatility and scalability.

[0207] S6. Use a meta-model to predict the heat of reaction.

[0208] S61. Input the previously acquired measurement signal, after preprocessing such as filtering and outlier removal, into the machine learning model. The model then outputs a predicted value of the heat release power during the process. Let this predicted power be P. 预 Test.

[0209] S62. The predicted heat release power P output by the machine learning model 预测 Substituting into the following formula, the heat released during the process, Q, can be calculated.

[0210]

[0211] Where V is the liquid volume (μL) in the microfluidic chip, representing the volume of liquid stored in the chip during titration or other testing processes; q is the flow rate of the material in the microfluidic channel (μL / min), which is usually set and monitored by a microfluidic pump or other precision flow rate control device; c is the concentration (mol / L) of the target substance in the test material, used for the subsequent calculation of the heat of reaction ΔH.

[0212] S63. Given that the target concentration of the tested material is c, the total feed amount or reactant content can be characterized under the condition that the microfluidic chip volume is V. To facilitate the calculation of energy-related chemical thermal effects, μL is usually converted to L.

[0213] S64. Based on the measured heat of release Q and concentration c, calculate the heat of reaction ΔH.

[0214]

[0215] S65. If the process is determined to be an endothermic process, then an additional exothermic power P needs to be provided to the thin-film heating element through a programmable DC power supply. 补偿 And ensure that the compensation value can cover the power absorbed during the heat absorption process, so as to ensure that the whole system is in the desired temperature or energy balance state.

[0216] S66, based on the predicted heat absorption power P and the compensation power P 补偿Determine the actual required heating element power:

[0217] P = P 补偿 -P 预测

[0218] Among them, P 预测 P represents the endothermic power predicted by the machine learning model. 补偿 It needs to be greater than |P 预测 |, to meet the energy compensation requirements of the external heat absorption process.

[0219] Whether the reaction is exothermic or endothermic, the exothermic power P can be calculated. 预测 Furthermore, a DC power supply compensation mechanism is employed to achieve dynamic balance, ensuring a more stable and accurate measurement process. Leveraging the predictive capabilities of machine learning models and the precise controllability of microfluidic systems in terms of flow rate and volume, the resolution and robustness of exothermic / endothermic heat measurements can be significantly improved, while simultaneously meeting the experimental requirements of various types of chemical systems or materials. This invention is not only applicable to titration calorimeter testing scenarios but can also be extended to other continuous flow reaction systems; it also exhibits good adaptability to different flow conditions, temperature, or pressure environments.

[0220] In order to correct the signal values, this invention uses a machine learning-based method to accurately calibrate the calorimeter.

[0221] Specifically, this invention controls the heating power of the thin-film heating element using an adjustable DC power supply to simulate the heat dissipation of the test section in microfluidic chip A. During the calibration phase, water is injected into both chips to obtain signal values ​​under different heat dissipation powers. First, as... Figure 6 As shown, this invention collects signal values ​​of the test section under different heat release powers, and according to... Figure 5The method preprocessed the data. Next, the model was initialized and parameters were set, creating a powerful gradient boosting regression model to handle nonlinear relationships in the data. Then, iterative training began. In each iteration, the data was divided into a training set and a test set, with a ratio of 80% and 20%, respectively. The model was trained and predictions were made on the training set, and residuals were calculated. To improve the model's robustness, the invention dynamically determined a threshold for outliers based on the standard deviation of the residuals, filtering out samples with absolute residuals greater than the threshold, obtaining a filtered training set, and retraining the model on it. Subsequently, the model was evaluated: predictions were made on the test set, the mean squared error (MSE) was calculated, and convergence was determined. If the current MSE was less than the set tolerance, the model was considered to have reached the expected accuracy, and iteration stopped; otherwise, the dataset was updated, and the next iteration began. Finally, the invention made predictions on the complete dataset, calculated the residuals and relative errors, and saved the final model. Through this method, the invention successfully achieved high-precision correction of calorimeter signals, significantly improving the model's reliability and prediction accuracy. Figure 7 The figure shows the error distribution of the calorimeter after calibration. It can be seen that in the test section, the error can be basically controlled below 5% when the heat release power is between 0.3-1000 microwatts, and the error is below 3% when the heat release power is between 30-1000 microwatts.

[0222] In summary, this invention can measure data that can be obtained using conventional isothermal titration calorimeters, and can be used to study the dynamic processes and thermodynamic parameters of intermolecular interactions, including but not limited to measuring the micelle concentration and micelle formation enthalpy of surfactants, protein-protein interactions, drug-DNA interactions, and protein-nucleic acid interactions. It can serve as an alternative research method to isothermal titration calorimeters.

[0223] Example 1

[0224] This example demonstrates the determination of micelle concentration and micelle formation enthalpy of the anionic surfactant sodium dodecylbenzenesulfonate (SDBS) using a continuous flow dynamic titration calorimetry method.

[0225] The first syringe pump A contains 0.01 mol / L CTAB, while the second syringe pump A, the first syringe pump B, and the second syringe pump B all contain water. The test temperature is set to 25℃. Table 1 shows the parameters set for each pump during the test and the corresponding target concentration of CTAB.

[0226] Table 1

[0227]

[0228] Based on the set pump parameters, the injection pump is simultaneously activated and continuously fed via the central control and data acquisition unit. Signals from thermoelectric generators A and B, as well as thermocouple signals, are continuously acquired via an analog-to-digital converter. Each pump operates for 10 minutes. Figure 5 Signal processing was performed. A machine learning model was used to predict the exothermic power of the test data. The liquid volume of the microfluidic chip was 154 μL. The heat of exothermic reaction and heat of reaction were calculated using Equations 1 and 2. The calculation results are shown in Table 2.

[0229] Table 2

[0230]

[0231]

[0232] Plot the data obtained. Figure 9 The Boltzmann function was used to fit the data, and Python was used to fit the data using the least squares method. The fitting process is as follows: Figure 8 As shown. The defined model function is:

[0233]

[0234] Among them, A max A is the maximum value of the heat of reaction. min Let be the minimum heat of reaction, and CMC be the micelle concentration. A was obtained through fitting. max =14.36±0.15kJ / mol, A min =2.00±0.10 kJ / mol, CMC=0.99±0.01 mmol. Calculate the enthalpy of micelle formation ΔH. mic ,

[0235] -ΔH mic =A max -A min =12.37 kJ / mol (5)

[0236] Example 2

[0237] This example demonstrates the determination of micelle concentration and micelle formation enthalpy of the anionic surfactant sodium dodecylbenzenesulfonate (SDBS) using a continuous flow dynamic titration calorimetry method.

[0238] The first syringe pump A contains 0.01 mol / L SDBS, while the second syringe pump A, the first syringe pump B, and the second syringe pump B all contain water. The test temperature was set to 25℃. Table 3 shows the parameters set for each pump during the test and the corresponding target concentration of SDBS.

[0239] Table 3

[0240]

[0241]

[0242] Based on the set pump parameters, the injection pump is simultaneously activated and continuously fed via the central control and data acquisition unit. Signals from thermoelectric generators A and B, as well as thermocouple signals, are continuously acquired via an analog-to-digital converter. Each pump operates for 10 minutes. Figure 5 Signal processing was performed. A machine learning model was used to predict the exothermic power of the test data. The liquid volume of the microfluidic chip was 154 μL. The heat of exothermic reaction and heat of reaction were calculated using Equations 1 and 2. The calculation results are shown in Table 4.

[0243] Table 4

[0244]

[0245] Plot the data obtained. Figure 10 The Boltzmann function is used to fit the data. The fitting process is as follows: Figure 8 As shown. The defined model function is Equation 3. A is obtained through fitting. max =16.19±0.55kJ / mol, A min =2.13±0.42kJ / mol, CMC=1.31±0.05mmol. Calculate the enthalpy of micelle formation ΔH. mic ,

[0246] -ΔH mic =A max -A min =14.06 kJ / mol (6)

[0247] Example 3

[0248] This example uses a continuous flow dynamic titration calorimetry method to test the micelle concentration and micelle formation enthalpy of the nonionic surfactant Triton X-100 (TX-100).

[0249] The first syringe pump A contains TX-100 at a concentration of 0.002 mol / L, while the second syringe pump A, the first syringe pump B, and the second syringe pump B all contain water. The test temperature was set to 25℃. Table 5 shows the parameters set for each pump during the test and the corresponding target concentration of TX-100.

[0250] Table 5

[0251]

[0252] The experiment revealed that the micelle formation process of TX-100 is endothermic. Using a programmable DC power supply, the heating power of the thin-film heating element was set to 5 microwatts. Therefore, the measured endothermic power P(TX-100) (microwatts) is...

[0253] P(TX-100)=5-P 预测 (7)

[0254] Where P 预测 The predicted value, in microwatts, is the heat release power of the test data predicted by a model built using machine learning.

[0255] Based on the set pump parameters, the injection pump is simultaneously activated and continuously fed via the central control and data acquisition unit. Signals from thermoelectric generators A and B, as well as thermocouple signals, are continuously acquired via an analog-to-digital converter. Each pump operates for 10 minutes. Figure 5 Signal processing was performed. A machine learning model was used to predict the exothermic power of the test data. The liquid volume of the microfluidic chip was 154 μL. The heat of exothermic reaction and heat of reaction were calculated using Equations 1 and 2. The calculation results are shown in Table 6.

[0256] Table 6

[0257]

[0258] Plot the data obtained. Figure 11 The Boltzmann function is used to fit the data. The fitting process is as follows: Figure 8 As shown. The defined model function is Equation 3. A is obtained through fitting. max = -2.70 ± 0.06 kJ / mol, A min = -8.40 ± 0.06 kJ / mol, CMC = 0.32 ± 0.00 mmol. Calculate the enthalpy of micelle formation ΔH. mic ,

[0259] ΔH mic =5.70 kJ / mol (8)

[0260] Comparative Example

[0261] The data obtained using this method are compared with data from the literature, as shown in Table 7.

[0262] Comparison with experimental data from different sources in the literature shows that the critical micelle concentration (CMC) and micelle formation enthalpy change (ΔHmic) measured using the method of this invention exhibit high consistency in terms of accuracy and repeatability. This invention utilizes continuous flow dynamic titration combined with microfluidics and machine learning correction, which can effectively reduce operational errors while shortening testing time, thereby achieving accurate detection of thermal effects.

[0263] Table 7 Comparison of Measured Values ​​and Literature Values

[0264]

[0265] Example 4

[0266] In this embodiment, the enzymatic reaction kinetics and thermal effects of trypsin and N-benzoyl-L-arginine ethyl ester hydrochloride (BAEE) were measured using a continuous flow dynamic titration calorimetry method.

[0267] Trypsin was dissolved in 0.05 mol / L phosphate-buffered saline (PBS) at pH 7.8 to prepare an enzyme solution with a concentration of 10 nmol / L. BAEE was dissolved in the same buffer to prepare an 800 μmol / L substrate solution. First syringe pump A, second syringe pump A, and first syringe pump B were filled with the above trypsin solution, while second syringe pump B was loaded with the substrate solution. The test temperature was set to 25°C, and all four syringe pumps were simultaneously activated via a central control and data acquisition unit for dynamic titration using a continuous feeding method. Table 8 shows a typical flow rate setting scheme in this embodiment. Each flow rate combination was run continuously for 10 minutes during the test to ensure flow path stability and to collect sufficient thermal signals.

[0268] Table 8

[0269]

[0270]

[0271] According to the set pump parameters, the software-controlled device continuously and thoroughly mixes the trypsin solution with BAEE. Utilizing the channel design of the microfluidic chip, the enzyme and substrate undergo an enzymatic reaction at a constant temperature of 25°C, accompanied by an exothermic effect. Similar to Example 1, the heat Q (mJ) generated or absorbed during the reaction is calculated, and the results are as follows: Figure 12 As shown.

[0272] Heat release Q obs The initial concentration [S] exhibits a linear relationship, assuming it satisfies the linear equation:

[0273] Q obs (i)=xΔH rxn [S] i V

[0274] Where x is the "common conversion ratio" achieved, and ΔH rxn The heat released per mole of substrate upon complete hydrolysis (J / mol), [S] i Let be the initial concentration for the i-th experiment, and V be the reaction volume. The slope ΔH is obtained by least squares fitting.rxn [S] i V. The final calculation yields ΔH. rxn It is -14.98 kJ / mol.

[0275] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A continuous flow dynamic titration calorimeter combining microfluidics and machine learning correction, characterized in that, include: The heat preservation device is used to place the calorimetric structure, keep the calorimetric structure warm, collect the ambient temperature of the calorimetric structure, and send it to the control and signal processing system. The calorimetric structure includes a material preheating aluminum block (461) and a constant temperature aluminum block (462) arranged on the left and right sides. A material inlet / outlet interface block A (451) is provided in the front groove of the material preheating aluminum block (461), and a material inlet / outlet interface block B (452) is provided in the rear groove of the material preheating aluminum block (461). A thermoelectric generator A (441) is provided at the front of the constant temperature aluminum block (462), and a thermoelectric generator B (442) is provided at the rear of the constant temperature aluminum block (462). A thin film heating element (43) is provided above the thermoelectric generator A (441), a microfluidic chip A (421) is provided above the thin film heating element (43), and a microfluidic chip B (422) is provided above the thermoelectric generator B (442). The rear of the constant temperature aluminum block (462) is connected to a constant temperature oil inlet (463), and the front of the constant temperature aluminum block (462) is connected to a constant temperature oil outlet (464). The lower part of the material inlet / outlet interface block A (451) is connected to the material inlet A1 (471), the material inlet A2 (472) and the material outlet A (473) respectively. The lower part of the material inlet / outlet interface block B (452) is connected to the material inlet B1 (474), the material inlet B2 (475) and the material outlet B (476) respectively. The first injection pump A (481), the second injection pump A (482), and the waste liquid tank (49) are connected to the microfluidic chip A (421) after passing through the material preheating aluminum block (461) and the material inlet / outlet interface block A (451); the first injection pump B (483), the second injection pump B (484), and the waste liquid tank (49) are connected to the microfluidic chip B (422) after passing through the material preheating aluminum block (461) and the material inlet / outlet interface block B (452); A control and signal processing system is used to control the programmable DC power supply (55) and the syringe pump controller (56); to collect signals generated by the thermoelectric generator and output calorimetric results; The control and signal processing system includes a central control and data acquisition system (51), signal amplifier A (521), signal amplifier B (522), low-pass filter A (531), low-pass filter B (532), analog-to-digital converter (54), programmable DC power supply (55), and injection pump controller (56). The positive and negative electrodes of the programmable DC power supply (55) are connected to the thin-film heating element (43) via wires. The signal generated by the thermoelectric generator A (441) is connected to the signal amplifier A (521), the low-pass filter A (531), and the analog-to-digital converter (54) in sequence via the signal transmission line. The signal generated by the thermoelectric generator B (442) is connected to the signal amplifier B (522), the low-pass filter B (532), and the analog-to-digital converter (54) in sequence via the signal transmission line. The analog-to-digital converter (54) transmits the signal to the central control and data acquisition system (51) via the signal transmission line. The central control and data acquisition system (51) controls the programmable DC power supply (55) and the injection pump controller (56) via the communication line. The method of using the continuous flow dynamic titration calorimeter that combines microfluidics and machine learning correction includes the following steps: S1. Materials are continuously added simultaneously to microfluidic chip A and microfluidic chip B at the same flow rate; wherein, in the test section, the material continuously introduced into microfluidic chip A through the first injection pump A is the titrant for the test titration process, and the material continuously introduced into microchannel reference chip A through the second injection pump A is the titrant for the test titration process; in the reference section, the materials continuously introduced into microfluidic chip B through the first injection pump B and the second injection pump B are both titrants; wherein the flow rates of the first injection pump A and the first injection pump B are the same, and the flow rates of the second injection pump A and the second injection pump B are the same. S2. The target concentration of the solution in microfluidic chip A is controlled by controlling the flow rate of each injection pump; the titration process is simulated by adjusting the flow rate ratio of the injection pumps; constant temperature oil is introduced into the material preheating aluminum block and the constant temperature aluminum block through the constant temperature oil inlet, and the temperature of the constant temperature oil is the set test temperature; the signal values ​​generated by thermoelectric generator A and thermoelectric generator B are amplified by 100 times by signal amplifier A and signal amplifier B respectively, and then high frequency noise is filtered out by low-pass filter A and low-pass filter B respectively. S3. Collect the signal values ​​of the two channels through an analog-to-digital converter and transmit them to the central control and data acquisition system; S4. The central control and data acquisition system collects the temperature signals generated by thermocouples, cleans the signal data, and obtains a valid dataset. S41. Under the same test conditions, continuously acquire the temperature signal generated by the thermocouple through the central control and data acquisition system for 10 minutes, and store the acquired temperature signal in the database, denoted as... ;in, ( ) indicates the first Temperature signal values ​​at each sampling time, This represents the total number of sampling points; S42. Perform a Fast Fourier Transform on the acquired temperature signal data to obtain the energy distribution of the signal in the frequency domain, and then convert the time domain signal { ( Transforming this to the frequency domain, it can be expressed as: Where ω is the frequency sampling point index, 0≤ < ; Removing DC components: soon Set the component corresponding to =0 to 0; find the maximum peak value: in Within the range >0, find the frequency component with the largest amplitude. max The corresponding frequency is obtained by conversion. ,in Indicates the signal sampling frequency; Set the cutoff frequency: 1.5× max The cutoff frequency of the Butterworth low-pass filter is denoted as . cut =1.5 max ; S43. Use a Butterworth low-pass filter to process the temperature signal. ( To remove high-frequency noise, a filter is applied. The transfer function of the filter can be expressed as: in, =2 cut Given the filter's angular cutoff frequency, the transfer function is discretized and mapped to... - Domain, thereby realizing digital filtering; the filtered data is denoted as { flt1 ( )}, thus obtaining preliminary filtering results; S44, Based on preliminary filtering results { flt1 ( Furthermore, a two-dimensional Kalman filter is used to estimate and correct the measurement signal; the system state model is as follows: = 1+ + in, Indicates the first The state vector at time t, This is an optional control vector. For process noise, and These are the system state transition matrix and control matrix, respectively; the observation model is: = + in, For the observed values, To observe the noise, The observation matrix is ​​used; in the Kalman filter recursive equation, the prediction stage is as follows: in, | 1 represents the prediction of state covariance. The process noise covariance matrix is ​​updated in the following stages: = | 1 ( | 1 + ) 1 | =( ) | 1 in, Here is the Kalman gain matrix. To observe the noise covariance matrix, the final temperature signal after two-dimensional Kalman filtering correction was obtained. flt2 ( )}; S45. Record the heat dissipation power data of the thin-film heating element as follows: The data is then unfolded into a one-dimensional data sequence in chronological order to maintain the comparability of heat release power at different time points. ( Sampling time of temperature signal flt2 ( One-to-one correspondence; S46. Use window functions to apply { ( Perform segmented analysis; set window width. =60, window movement step size =2; for the first The power data within each window is statistically analyzed, and the data within each window is { ( ) | ∈[ start , end ]}, calculate the first-order quartiles within this window. 1 and third-order quartiles 3, of which 1 represents the value at the 25th percentile after sorting the data in ascending order. 3 represents the value at the 75th percentile after sorting in ascending order; S47. For the data in each window { ( Perform the following judgments to calculate the interquartile range (IQR) within the window: IQR= 3 1 Determine if the data is an outlier. ( ) Exceeding the range [ 1 1.5×IQR, If the value is greater than or equal to 3 + 1.5 × IQR, it is considered an outlier; all heat release power values ​​marked as outliers and their corresponding measurement signal data are considered outliers. flt2 ( All data points are discarded, retaining only normal data and their temperature signal pairs; through the above window function traversal and anomaly marking, extreme noise points occurring during observation or acquisition are removed, thereby obtaining a valid dataset {( flt2 ( ), ( ))}; S5. Employ machine learning-based methods to build a meta-model based on an effective dataset; S51. Three algorithms—XGBoost, LightGBM, and Random Forest—are used to model and train the training data, respectively. Parameters are then tuned using grid search, random search, or Bayesian optimization to obtain relatively stable parameter configurations, thus forming three sets of base models. The objective function of XGBoost is: in, Indicates the true label, For the cumulative prediction of the first m-1 trees, Indicates the first m The residual function learned by the tree This is a regularization term for tree complexity. For the The output of the LightGBM model can be approximated as: in, For the first The learning rate of each tree. Indicates the first Tree pairs of input The estimation; Random forest is constructed by The final prediction output is obtained by voting on or averaging CART decision trees. The model can be represented as follows: in, For the first The output of each decision tree; S52. After tuning the hyperparameters of XGBoost, LightGBM, and Random Forest respectively, the optimal combination of hyperparameters was finally locked and used to train three base models, denoted as: S53. Input the same training or validation set data into the three base models in sequence to obtain three sets of predicted values: in, X Represents the input feature matrix. , , Let XGBoost, LightGBM, and Random Forest represent the predicted heat release vectors, respectively. Then, the predicted heat release vectors are compared with their corresponding actual values. By comparing the results, we obtained the error and residual information generated by the three models. S54, Real Value As training labels for the meta-model, and , , The new feature vector is formed by concatenating or combining the features, denoted as: S55. Support vector machine regression is used as the meta-model, denoted as... SVM Its objective function can be adopted. -Insensitive loss form: in, and These are the weight vector and bias term of the support vector machine, respectively. The regularization coefficient is . Indicates exceeding - Slack variables in insensitive intervals; kernel function RBF kernel, linear kernel or polynomial kernel can be selected according to actual needs; S56. Use cross-validation, hold-out validation, or other validation techniques to adjust and optimize the parameters of the SVM, including the kernel function type and regularization coefficient. Kernel function parameters Specifically, it can be expressed as: Here, CVBloss represents the comprehensive loss value obtained by measuring and averaging the prediction error during cross-validation; through this process, the meta-model fully learns the mapping relationship between the base model output and the true label; S57. When the prediction error of the meta-model on the validation set meets the preset accuracy requirements, the parameter determination of the meta-model is completed, and the final meta-model is obtained. ; S6. Use a meta-model to predict the heat of reaction.

2. The continuous flow dynamic titration calorimeter combining microfluidics and machine learning correction according to claim 1, characterized in that, The heat preservation device includes a calorimeter outer shell cover (21), a calorimeter outer shell base (22), and a thermocouple (3). The calorimeter outer shell cover (21) and calorimeter outer shell base (22) are located outside the calorimeter structure, the thermal insulation felt cover (11) is located outside the calorimeter outer shell cover (21), the thermal insulation felt base (12) is located outside the calorimeter outer shell base (22), and the thermocouple (3) is located inside the calorimeter outer shell cover (21).

3. The continuous flow dynamic titration calorimeter combining microfluidics and machine learning correction according to claim 1, characterized in that, Silicone sealing rings (453) are provided on the material inlet / outlet interface block A (451) and the material inlet / outlet interface block B (452).

4. The continuous flow dynamic titration calorimeter combining microfluidics and machine learning correction according to claim 1, characterized in that, Material in / out interface block A (451) and material in / out interface block B (452) are provided with a material in / out viewing window (411), and microfluidic chip A (421) and microfluidic chip B (422) are provided with a microfluidic viewing window (412).

5. The continuous flow dynamic titration calorimeter combining microfluidics and machine learning correction according to claim 1, characterized in that, The microfluidic chip A (421) and microfluidic chip B (422) have the same structure, and the thermoelectric generator A (441) and thermoelectric generator B (442) are of the same model.

6. The continuous flow dynamic titration calorimeter combining microfluidics and machine learning correction according to claim 1, characterized in that, When the actual ambient temperature exceeds the temperature range of the machine learning training set, a method is used to improve the robustness of the model output in extrapolation scenarios through linear mapping to map the temperature outside the temperature range to the temperature range. The steps are as follows: P1, Temperature based on existing training dataset With measurement signal The correspondence between them can be used to obtain a linear fitting model using the least squares method or other linear regression methods, denoted as: = + in, Indicates the measurement signal. and Here, the slope and intercept of the linear model are given by the following formulas: in, and For the training set Temperature and measurement signals of each sample, and These are the average values ​​of the temperature and the measured signal, respectively. P2. Set the temperature range for the training data as [ T min , T max This interval can be considered as the core region where the model has high predictive reliability in the temperature dimension; { , }as well as T min , T max Stored in the parameter library for later use; P3. For the new test dataset ,in For the first j Temperature values ​​of each test sample Other feature vectors or measurement signals representing this sample need to be determined first. Does it fall into [ T min , T max Within the range; if If the temperature is normal, no correction is needed; if or If the temperature exceeds the range, it is marked as a sample and added to the set. middle; P4. For samples marked as out of range The linear relationship is invoked to correct the corresponding measurement signal; assuming the original value of the measurement signal for out-of-range samples is... Then the mapping correction formula is: in, This indicates the corrected measurement signal, in conjunction with... Ensure new test samples Closer to the training range T min , T max The corresponding feature distribution; P5. Replace all the original out-of-range data with the linearly corrected test samples to form a new test dataset. At this point, the existing meta-model can be used to predict the updated test data.

7. The continuous flow dynamic titration calorimeter combining microfluidics and machine learning correction according to claim 1, characterized in that, The specific steps for S6 are as follows: S61. Determine whether the process is endothermic or exothermic. If the process is exothermic, execute S62~S65; if the process is exothermic, execute S66, S67, S63, S64, and S65 in sequence. S62. Input the preprocessed measurement signal, which has been collected in the early stage and filtered and removed outliers, into the machine learning model. The meta-model outputs the predicted value of the heat release power of the process, and this predicted value is denoted as S62. 预测 = P ; S63. The predicted heat release power output by the machine learning model. Substituting into the following formula, the heat released during the process can be calculated. : in, V The liquid volume within the microfluidic chip represents the volume of liquid stored in the chip during titration or other testing processes. q Δc represents the flow rate of the material in the microfluidic channel, typically set and monitored using a microfluidic pump or other precision flow rate control device; c represents the concentration of the target substance in the test material, used for subsequent reaction heat Δt. Calculation; S64. The target concentration of the tested material is known to be... Then, in the microfluidic chip volume is Under these conditions, the total feed amount or reactant content can be characterized; Convert to L; S65. Based on the measured heat release With concentration The heat of reaction Δ was calculated. : S66. If the process is determined to be an endothermic process, then additional exothermic power needs to be provided to the thin-film heating element through a programmable DC power supply. 补偿 And ensure heat dissipation power 补偿 It can cover the power absorbed during the heat absorption process, ensuring that the entire system is in the desired temperature or energy balance state; S67, based on the predicted heat absorption power and compensation power 补偿 Determine the actual required heating element power: P=P 补偿 -P 预测 Among them, P 预测 P represents the endothermic power predicted by the machine learning model. 补偿 Must be greater than | 预测 |, to meet the energy compensation requirements of the external heat absorption process.

Images