An infrared transparent conductive material sub-band screening method based on machine learning and multi-objective optimization

Abstract

The application discloses an infrared transparent conductive material waveband screening method based on machine learning and multi-objective optimization, and aims to solve the problems that the existing evaluation method is difficult to represent the free carrier plasma reflection edge, and the transparency and the conductivity are restricted to each other, and it is difficult to evaluate by a single performance index. The screening method comprises the following steps: taking a material descriptor as an input, respectively establishing a band gap prediction model, a refractive index prediction model and an electron transport parameter prediction model; using the three prediction models to obtain the predicted band gap, the predicted refractive index and the predicted electron transport parameter of a candidate material; determining a high-frequency dielectric parameter according to the predicted refractive index or dielectric property, and calculating the free carrier plasma reflection edge; establishing a waveband candidate set; and performing Pareto non-dominated sorting on the transparency evaluation quantity and the conductivity evaluation quantity of the candidate material. The application can quickly form independent candidate sets of short-wave, medium-wave and long-wave infrared in a large-scale material space, and the omission can be avoided.

Description

Technical Field

[0001] This invention belongs to the fields of materials informatics, infrared optoelectronic materials, and computer-aided materials screening technology. Specifically, it relates to a method for screening infrared transparent conductive materials by using material descriptors and machine learning models to predict material band gap, refractive index, and electron transport parameters, and combining free carrier plasma reflection edge, band-specific constraints, and multi-objective optimization. Background Technology

[0002] Infrared transparent conductive materials need to possess a certain degree of optical transmittance in the target infrared band, while also having the carrier transport capability required for the application. They can be used for infrared detection, infrared windows, transparent electrodes, and related optoelectronic devices. The performance of infrared transparent conductive materials is influenced by a variety of factors, including the material's band gap, refractive index, dielectric response, free carrier concentration, effective carrier mass, and carrier scattering. Obtaining the complete optoelectronic properties of a large number of candidate materials one by one through experiments or first-principles calculations requires significant time and computational resources, making it difficult to directly use for rapid screening in large-scale material spaces.

[0003] The band gap can be used to determine whether interband absorption may occur within the target wavelength band; the refractive index is related to the dielectric response and interface reflection of the material; and electron transport parameters can reflect the transport capacity of free carriers in the material. Because the data sources, sample sizes, and numerical distributions of the above properties differ, obtaining multiple properties simultaneously using the same prediction model may affect the prediction stability.

[0004] Existing evaluation methods for transparent conductive materials typically use one or more of the following indicators as screening criteria: band gap, conductivity, or transmittance. However, infrared transparency windows are also affected by the reflection of free carrier plasma, making it difficult to directly determine whether the reflected free carriers have entered the target infrared band using only band gap and conductivity.

[0005] Increasing the free carrier concentration generally improves the electrical transport capability of materials, but it also increases the characteristic frequency of free carrier plasma, causing the plasma reflection edge to shift towards shorter wavelengths. When the plasma reflection edge enters the target infrared band, reflection within the target band may be enhanced, thus limiting the infrared transparency of the material.

[0006] Furthermore, there is a trade-off between transparency and conductivity. Setting a single threshold based solely on one property or ranking materials from high to low may exclude candidate materials that excel in another property.

[0007] Therefore, there is a need for a screening method for infrared transparent conductive materials that combines machine learning prediction of material properties, calculation of free carrier plasma reflection edges, band-specific screening, and Pareto multi-objective optimization. Summary of the Invention

[0008] This invention aims to address the problems of high costs associated with obtaining candidate material band gap, refractive index, and electron transport parameters one by one, the difficulty of characterizing the free carrier plasma reflection edge in existing evaluation methods, and the mutual constraints between transparency and conductivity, making it difficult to evaluate by a single performance index. Instead, it provides a band-based screening method for infrared transparent conductive materials based on machine learning and multi-objective optimization.

[0009] The present invention, based on machine learning and multi-objective optimization, provides a method for screening infrared transparent conductive materials by wavelength band, which is implemented according to the following steps:

[0010] Step 1: Combine Magpie elemental property statistics, different stoichiometric norms, average s, p, d and f valence electron counts, material ionicity feasibility markers, maximum ionicity and average ionicity into a material descriptor. Use the material descriptor as input and the material band gap as the prediction target to train the band gap prediction model and obtain the trained band gap prediction model.

[0011] Step 2: Using the material descriptor as input and the material refractive index as the prediction target, train the refractive index prediction model to obtain the trained refractive index prediction model.

[0012] Step 3: Using the material descriptor as input and the material electronic transport parameter S as the prediction target, train the electronic transport parameter prediction model to obtain the trained electronic transport parameter prediction model.

[0013] Step 4: Convert the material to be screened into the corresponding material descriptor and input it into the trained bandgap prediction model, the trained refractive index prediction model and the trained electron transport parameter prediction model respectively to obtain the predicted bandgap Eg, the predicted refractive index n and the predicted electron transport parameter S respectively.

[0014] Step 5: Obtain the high-frequency relative permittivity ε from the predicted refractive index n. ∞ ;

[0015] When carrier damping is neglected, the angular frequency of free carrier plasma reflection is expressed as:

[0016]

[0017] When carrier damping is considered, the angular frequency of free carrier plasma reflection is expressed as:

[0018]

[0019] Where ε0 is the vacuum permittivity and c is the speed of light in vacuum. The carrier damping rate;

[0020] The plasma reflection edge wavelength is determined according to the following formula:

[0021]

[0022] in, The vacuum permittivity, It is the speed of light in a vacuum;

[0023] Step 6: Set the target infrared band B, which is represented as follows:

[0024]

[0025] in, The shortest wavelength in the target band. The longest wavelength in the target band;

[0026] Establish preliminary constraints on interband absorption and free carrier plasma reflection edges;

[0027] A set of candidate materials was obtained through preliminary screening using interband absorption confinement and free carrier plasma reflection edge confinement.

[0028] Step 7: Set target wavelength ranges for short-wave infrared, mid-wave infrared and long-wave infrared respectively, and calculate the transparency evaluation value and conductivity evaluation value of each candidate material in the candidate material set in the corresponding target infrared band B.

[0029] The transparency evaluation parameters are bandgap constraint margin, plasma reflection edge constraint margin, or single-interface reflectivity.

[0030] The bandgap constraint margin Represented as:

[0031] ;

[0032] The plasma reflection edge constraint margin Represented as:

[0033] ;

[0034] The single interface reflectivity Represented as:

[0035] ;

[0036] The conductivity evaluation quantity is the predicted electron transport parameter S, and n is the predicted refractive index;

[0037] Step 8: Using transparency and conductivity as optimization objectives, perform Pareto non-dominated sorting on the candidate material set to obtain the Pareto frontier materials;

[0038] Step 9: Calculate the infrared comprehensive evaluation parameter IR-FOM for Pareto front materials. B And according to the infrared comprehensive evaluation parameter IR-FOM B Secondary sorting and infrared comprehensive evaluation parameters IR-FOM B Represented as:

[0039]

[0040] in, For the target infrared band The corresponding transparency evaluation metric, For the target infrared band The corresponding conductivity evaluation quantity, An evaluation function that simultaneously considers both transparency and conductivity evaluation metrics;

[0041] After secondary sorting, the selected infrared transparent conductive materials are obtained, thus completing the band-based screening method for infrared transparent conductive materials based on machine learning and multi-objective optimization.

[0042] The infrared transparent conductive material described in this invention refers to an inorganic crystalline material with infrared transmission potential and conductivity potential. Specifically, it refers to inorganic elements or inorganic compounds with a definite chemical composition and crystal structure, including but not limited to oxides, sulfides, selenides, tellurides, nitrides, phosphides, halides, and other inorganic crystalline materials. The screening targets of this invention do not include small organic molecule materials, conductive polymers, organic-inorganic composite materials, and amorphous materials without a definite crystal structure.

[0043] This invention utilizes a unified and reproducible material descriptor to train prediction models for bandgap, refractive index, and electron transport parameters, and combines the model outputs with Drude free carrier plasma reflection edge, band-specific interband absorption constraints, and Pareto multi-objective optimization. This enables the rapid formation of independent candidate sets for short-wave, mid-wave, and long-wave infrared radiation in a large-scale material space, avoiding the omission of preferred materials caused by sorting based solely on a single optical or electrical index. Simultaneously, it outputs each predicted quantity, constraint score, Pareto label, and IR-FOM, facilitating result traceability, verification, and subsequent first-principles calculations and sample experimental validation, thereby reducing the computational and experimental costs of the initial screening process. Attached Figure Description

[0044] Figure 1 This is the overall flowchart of the machine learning band-based screening of this invention;

[0045] Figure 2 This is a diagram showing the relationship between the plasma edge and the target waveband;

[0046] Figure 3This is a comparison chart of the prediction performance of XGBoost, Random Forest, and Multilayer Perceptron in the embodiments;

[0047] Figure 4 The above are the predicted results of band gap, refractive index, and electronic transport parameters, along with a feature importance graph, in the examples.

[0048] Figure 5 This is a diagram showing the distribution, number of candidates, and overall evaluation ranking of candidate materials in different infrared bands in the embodiments. Figure 5 (a) is a distribution diagram of feasible materials in three infrared bands and the Pareto front on the infrared transparency-conductivity score plane; Figure 5 (b) Distribution map of Pareto front materials and their material identifiers in three infrared bands; Figure 5 (c) is a comparison chart of the number of feasible materials in the three infrared bands; Figure 5 (d) is the IR-FOM ranking diagram of Pareto front materials in three infrared bands. Detailed Implementation

[0049] Specific Implementation Method 1: This implementation method, based on machine learning and multi-objective optimization, employs a band-specific screening method for infrared transparent conductive materials, and is carried out according to the following steps:

[0050] Step 1: Combine Magpie elemental property statistics, different stoichiometric norms, average s, p, d and f valence electron counts, material ionicity feasibility markers, maximum ionicity and average ionicity into a material descriptor. Use the material descriptor as input and the material band gap as the prediction target to train the band gap prediction model and obtain the trained band gap prediction model.

[0051] Step 2: Using the material descriptor as input and the material refractive index as the prediction target, train the refractive index prediction model to obtain the trained refractive index prediction model.

[0052] Step 3: Using the material descriptor as input and the material electronic transport parameter S as the prediction target, train the electronic transport parameter prediction model to obtain the trained electronic transport parameter prediction model.

[0053] Step 4: Convert the material to be screened into the corresponding material descriptor and input it into the trained bandgap prediction model, the trained refractive index prediction model and the trained electron transport parameter prediction model respectively to obtain the predicted bandgap Eg, the predicted refractive index n and the predicted electron transport parameter S respectively.

[0054] Step 5: Obtain the high-frequency relative permittivity ε from the predicted refractive index n. ∞ ;

[0055] When carrier damping is neglected, the angular frequency of free carrier plasma reflection is expressed as:

[0056]

[0057] When carrier damping is considered, the angular frequency of free carrier plasma reflection is expressed as:

[0058]

[0059] Where ε0 is the vacuum permittivity and c is the speed of light in vacuum. The carrier damping rate;

[0060] The plasma reflection edge wavelength is determined according to the following formula:

[0061]

[0062] in, The vacuum permittivity, It is the speed of light in a vacuum;

[0063] Step 6: Set the target infrared band B, which is represented as follows:

[0064]

[0065] in, The shortest wavelength in the target band. The longest wavelength in the target band;

[0066] Establish preliminary constraints on interband absorption and free carrier plasma reflection edges;

[0067] A set of candidate materials was obtained through preliminary screening using interband absorption confinement and free carrier plasma reflection edge confinement.

[0068] Step 7: Set target wavelength ranges for short-wave infrared, mid-wave infrared and long-wave infrared respectively, and calculate the transparency evaluation value and conductivity evaluation value of each candidate material in the candidate material set in the corresponding target infrared band B.

[0069] The transparency evaluation parameters are bandgap constraint margin, plasma reflection edge constraint margin, or single-interface reflectivity.

[0070] The bandgap constraint margin Represented as:

[0071] ;

[0072] The plasma reflection edge constraint margin Represented as:

[0073] ;

[0074] The single interface reflectivity Represented as:

[0075] ;

[0076] The conductivity evaluation quantity is the predicted electron transport parameter S, and n is the predicted refractive index;

[0077] Step 8: Using transparency and conductivity as optimization objectives, perform Pareto non-dominated sorting on the candidate material set to obtain the Pareto frontier materials;

[0078] Step 9: Calculate the infrared comprehensive evaluation parameter IR-FOM for Pareto front materials. B And according to the infrared comprehensive evaluation parameter IR-FOM B Secondary sorting and infrared comprehensive evaluation parameters IR-FOM B Represented as:

[0079]

[0080] in, For the target infrared band The corresponding transparency evaluation metric, For the target infrared band The corresponding conductivity evaluation quantity, An evaluation function that simultaneously considers both transparency and conductivity evaluation metrics;

[0081] After secondary sorting, the selected infrared transparent conductive materials are obtained, thus completing the band-based screening method for infrared transparent conductive materials based on machine learning and multi-objective optimization.

[0082] In this embodiment, the predicted bandgap Eg is used to establish the short-wavelength cutoff of the transparent window of the material corresponding to the interband absorption constraint of the target band and to calculate the bandgap constraint margin; the predicted refractive index n is used to describe the transmittance of the transparent window of the material; and the predicted electron transport parameter S is used to characterize the intrinsic transport capability of the material and the long-wavelength cutoff of the transparent window and to calculate the plasma reflection edge constraint margin.

[0083] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that the bandgap prediction model in step one is an XGBoost regression model or a random forest model.

[0084] In this embodiment, the preferred bandgap prediction model is the XGBoost regression model.

[0085] Specific Implementation Method 3: This implementation method differs from Specific Implementation Method 1 or 2 in that the refractive index prediction model in step 2 is an XGBoost regression model or a random forest model.

[0086] In this embodiment, the preferred refractive index prediction model is the XGBoost regression model.

[0087] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that the electron transport parameter prediction model in step three is an XGBoost regression model or a random forest model.

[0088] In this embodiment, the preferred electronic transport parameter prediction model is the XGBoost regression model.

[0089] Specific Implementation Method 5: This implementation method differs from Specific Implementation Methods 1 to 4 in that when the bandgap prediction model, refractive index prediction model, and electron transport parameter prediction model all use the XGBoost regression model, the number of regression trees is set to 400, and the maximum depth of a single tree is set to 5.

[0090] Specific Implementation Method Six: This implementation method differs from Specific Implementation Method Five in that the learning rate in the XGBoost regression model is set to 0.05.

[0091] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that the electron transport parameter S in step three is conductivity. Ratio to carrier relaxation time τ:

[0092] .

[0093] In this embodiment, the material conductivity is expressed as follows in the Drude free carrier model:

[0094]

[0095] therefore:

[0096]

[0097] Where N is the free carrier concentration, e is the elementary charge, and m* is the effective carrier mass.

[0098] In this embodiment, when the values ​​of electronic transport parameters span multiple orders of magnitude, a logarithmic transformation can be performed after unifying the units. For example, we can let:

[0099]

[0100] Wherein, S0 is a preset reference quantity with the same unit as S, used to make the input of the logarithmic function dimensionless.

[0101] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One through Seven in that the predicted refractive index and the high-frequency relative permittivity in step five satisfy the following:

[0102]

[0103] Where, ε ∞ is the high-frequency relative permittivity, and n is the predicted refractive index.

[0104] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight in that the expression for establishing the inter-band absorption constraint in step six is ​​as follows:

[0105]

[0106] Where h is Planck's constant and c is the speed of light in a vacuum.

[0107] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One through Nine in that step six establishes the initial constraint on the free carrier plasma reflection edge.

[0108] .

[0109] Specific Implementation Method Eleven: This implementation method differs from Specific Implementation Methods One through Nine in that the infrared comprehensive evaluation parameter IR-FOM is used in step Nine. B Represented as:

[0110] .

[0111] Example: This example demonstrates a band-based screening method for infrared transparent conductive materials based on machine learning and multi-objective optimization, implemented according to the following steps:

[0112] Step 1: Combine the Magpie elemental property statistics, stoichiometric norms (0, 2, 3, 5, 7, and 10), average number of s, p, d, and f valence electrons, material ionic feasibility markers, maximum ionicity, and average ionicity into a material descriptor. The Magpie elemental property statistics include 22 fundamental Magpie elemental properties: atomic number, Mendeleev number, atomic mass, melting point, periodic table column and row, covalent radius, electronegativity, number of s / p / d / f valence electrons, total number of valence electrons, number of unfilled s / p / d / f orbitals, total number of unfilled orbitals, ground-state atomic volume, ground-state band gap, ground-state magnetic moment, and space group number. For each element in the material to be screened, calculate the minimum, maximum, range, composition-weighted average, mean absolute deviation, and mode of the fundamental Magpie elemental properties. Among them, the minimum value is the smallest of the basic attribute values ​​of each component element, the maximum value is the largest of them, the range is the difference between the maximum and minimum values, the weighted average value is the average value calculated by weighting the atomic fractions of each component element, the mean absolute deviation is the average value calculated by weighting the absolute deviations of the attribute values ​​of each component element relative to the weighted average value by atomic fractions, and the mode is the attribute value corresponding to the component element with the largest atomic fraction. Thus, a 22×6=132-dimensional Magpie element attribute statistic is formed.

[0113] The stoichiometric norm is the mole fraction of each element after normalization to the chemical formula. Based on this: the 0th-order norm represents the number of distinct elements in the composition; when p is 2, 3, 5, 7, or 10, the p-order stoichiometric norm is... The average number of s, p, d, and f valence electrons is the weighted average of the corresponding valence electron numbers of each element according to the mole fraction of the composition; the feasibility label for material ionicity indicates whether a combination satisfying charge neutrality can be found based on common oxidation states; if found, it is marked as True, otherwise it is marked as False; the maximum ionicity and average ionicity are both dimensionless ionicity indices.

[0114] Taking LiSbS(mp-756006) as an example, its stoichiometric norms of the 0th, 2nd, 3rd, 5th, 7th and 10th orders are 3, 0.577350, 0.480750, 0.415244, 0.389977 and 0.372041, respectively; the average number of s, p, d and f valence electrons are 1.666667, 2.333333, 3.333333 and 0, respectively; the material's ionicity is marked as True, the maximum ionicity is 0.472708 and the average ionicity is 0.087714.

[0115] Taking ZnSnN2 (mp-1029469) as an example, its stoichiometric norms of the 0th, 2nd, 3rd, 5th, 7th and 10th orders are 3, 0.612372, 0.538609, 0.506099, 0.501109 and 0.500098 respectively; the average number of s, p, d and f valence electrons are 2, 2, 5 and 0 respectively; the material ionicity is marked as True, the maximum ionicity is 0.383086 and the average ionicity is 0.080986.

[0116] Taking YSbPt(mp-4964) as an example, its stoichiometric norms of the 0th, 2nd, 3rd, 5th, 7th and 10th orders are 3, 0.577350, 0.480750, 0.415244, 0.389977 and 0.372041, respectively; the average number of s, p, d and f valence electrons are 1.666667, 1, 6.666667 and 4.666667, respectively; the material's ionicity is marked as False, the maximum ionicity is 0.244896 and the average ionicity is 0.046249.

[0117] Using the material descriptor as input and the material bandgap as the prediction target, the bandgap prediction model is trained to obtain the trained bandgap prediction model.

[0118] Step 2: Using the material descriptor as input and the material refractive index as the prediction target, train the refractive index prediction model to obtain the trained refractive index prediction model.

[0119] Step 3: Using the material descriptor as input and the material electronic transport parameter S as the prediction target, train the electronic transport parameter prediction model to obtain the trained electronic transport parameter prediction model.

[0120] Step 4: Convert the material to be screened into the corresponding material descriptor and input it into the trained bandgap prediction model, the trained refractive index prediction model and the trained electron transport parameter prediction model respectively to obtain the predicted bandgap Eg, the predicted refractive index n and the predicted electron transport parameter S respectively.

[0121] Step 5: Obtain the high-frequency relative permittivity ε from the predicted refractive index n. ∞， The predicted refractive index and the high-frequency relative permittivity satisfy the following:

[0122]

[0123] Where, ε ∞ is the high-frequency relative permittivity, and n is the predicted refractive index;

[0124] When carrier damping is neglected, the angular frequency of free carrier plasma reflection is expressed as:

[0125]

[0126] When carrier damping is considered, the angular frequency of free carrier plasma reflection is expressed as:

[0127]

[0128] Where ε0 is the vacuum permittivity and c is the speed of light in vacuum. The carrier damping rate;

[0129] The plasma reflection edge wavelength is determined according to the following formula:

[0130]

[0131] in, The vacuum permittivity, It is the speed of light in a vacuum.

[0132] Step 6: Set the target infrared band B, which is represented as follows:

[0133]

[0134] in, The shortest wavelength in the target band. The longest wavelength in the target band;

[0135] The expression for the established interband absorption constraint is:

[0136]

[0137] Where h is Planck's constant and c is the speed of light in a vacuum.

[0138] Establish preliminary constraints on the free carrier plasma reflection edge:

[0139]

[0140] A set of candidate materials was obtained through preliminary screening using interband absorption confinement and free carrier plasma reflection edge confinement.

[0141] Step 7: Set target wavelength ranges for short-wave infrared, mid-wave infrared and long-wave infrared respectively, and calculate the transparency evaluation value and conductivity evaluation value of each candidate material in the candidate material set in the corresponding target infrared band B.

[0142] In this embodiment, the normalized plasma reflection edge constraint margin, i.e., the infrared transparency score ir_trans_score, is selected as the transparency evaluation metric in Pareto optimization; the bandgap constraint margin is normalized to form gap_score, which is used for feasibility threshold judgment; the single interface reflectivity is an optional transparency evaluation metric, but it is not used as the Pareto ranking target in this embodiment.

[0143] The bandgap constraint margin Represented as:

[0144] ;

[0145] The plasma reflection edge constraint margin Represented as:

[0146] ;

[0147] The conductivity evaluation metric is the predicted electron transport parameter S;

[0148] Step 8: Using transparency and conductivity as optimization objectives, perform Pareto non-dominated sorting on the candidate material set to obtain the Pareto frontier materials;

[0149] Step 9: Calculate the infrared comprehensive evaluation parameter IR-FOM for Pareto front materials. B And according to the infrared comprehensive evaluation parameter IR-FOM B Secondary sorting and infrared comprehensive evaluation parameters IR-FOM B Represented as:

[0150]

[0151] in, For the target infrared band The corresponding transparency evaluation metric, For the target infrared band The corresponding conductivity evaluation quantity, An evaluation function that simultaneously considers both transparency and conductivity evaluation metrics;

[0152] After secondary sorting, the selected infrared transparent conductive materials are obtained, thus completing the band-based screening method for infrared transparent conductive materials based on machine learning and multi-objective optimization.

[0153] In step one of this embodiment, material records with dielectric data are obtained through the Materials Project. Materials with band gaps of 0.1–6.0 eV, energies higher than the convex hull value by no more than 0.1 eV / atom, and 2–4 constituent elements are selected, while excluding Tc, Pm, Po, At, Rn, Ra, Ac, Th, Pa, U, and actinides. After deleting records with missing band gaps or refractive indices, the electron transport dataset contains 30,467 positive values. In this embodiment, 5000 records are randomly selected from a fixed seed of 0 for training the electron transport model. This embodiment uses the StrToComposition function of the Matminer software to convert chemical formulas into Composition objects.

[0154] In this embodiment, XGBoost, Random Forest, and Multilayer Perceptron are used to establish bandgap prediction models, refractive index prediction models, and electronic transport parameter prediction models, respectively, in steps one, two, and three. 80% of the samples are used as the training set, and 20% as the independent test set. Five-fold random cross-validation is used within the training set, with a random seed of 0. After model evaluation, the final model is retrained using all corresponding samples.

[0155] Algorithm Comparison: The training / test split was kept identical, and the model was initialized with random seeds 0-4, repeated 5 times. The average R-value of XGBoost, Random Forest, and Multilayer Perceptron on the test set in bandgap tasks was [not specified]. 2 The targets are 0.816, 0.814, and 0.781, respectively; the targets for refractive index are 0.650, 0.653, and 0.621, respectively; and the targets for electron transport are 0.545, 0.532, and 0.426, respectively. Figure 3 As shown, the tree model is superior to the multilayer perceptron baseline in this embodiment, so XGBoost is chosen to unify the three batch prediction processes.

[0156] Under the given dataset and parameter conditions, the overall prediction results of XGBoost and Random Forest are close and superior to the adopted multilayer perceptron baseline. To unify the batch inference process for the three performance tasks, this embodiment uses the XGBoost regression model to establish bandgap prediction, refractive index prediction, and electron transport parameter prediction models, respectively. The parameter settings for the XGBoost regression model are as follows: number of regression trees n_estimators = 400, maximum depth of a single tree max_depth = 5, learning rate learning_rate = 0.05, row sampling ratio subsample = 0.8, column sampling ratio colsample_bytree = 0.8, and random seed random_state = 0.

[0157] In this embodiment, 6007 samples were used for both the bandgap and refractive index models, with 4805 samples in the training set and 1202 samples in the test set. The R², MAE, and RMSE of the bandgap model training set after 5-fold cross-validation were 0.824, 0.443 eV, and 0.605 eV, respectively, while those on the independent test set were 0.815, 0.453 eV, and 0.618 eV. The R², MAE, and RMSE of the refractive index model training set after 5-fold cross-validation were... 2 The MAE and RMSE were 0.520, 0.199 and 0.719, respectively, and 0.659, 0.197 and 0.571, respectively, on the independent test set.

[0158] In this embodiment, the electronic transport model uses log10(σ / τ) as the prediction target and employs 5000 samples, with 4000 samples in the training set and 1000 samples in the test set. The training set undergoes 5-fold cross-validation with R0. 2 The MAE and RMSE were 0.515, 0.374, and 0.540, respectively, and 0.546, 0.362, and 0.505 for the independent test set. After prediction, the SI scale value used for plasma reflection edge calculation was recovered according to σ / τ=10^[log10(σ / τ)]×5.222×10^17.

[0159] Feature importance analysis: such as Figure 4 As shown, in the band gap model, the mode of melting point, the maximum number of d-valence electrons, and the mode of covalent radius are of high importance; in the refractive index model, the maximum electronegativity, the minimum melting point, and the range of p-valence electrons are of high importance; and in the electron transport model, the maximum number of unfilled d orbitals is of the highest importance. These importance indicators only represent the statistical contribution of the model to this dataset and do not indicate a definitive causal relationship.

[0160] Batch prediction and plasma reflection edge calculation: For 6007 records to be screened, Eg, n, and log10(σ / τ) are predicted according to the 145-dimensional feature order of the training phase, and the high-frequency relative permittivity is taken. Under the approximation condition of neglecting carrier damping, the following is adopted: (where ω) p (where ε0 is the angular frequency of free carrier plasma reflection and ε0 is the vacuum permittivity) and (Where c is the speed of light in vacuum) Calculate the plasma reflection edge wavelength. The calculated λp range is 1.271–101.342 μm, with a median of 2.9999 μm.

[0161] Bandwidth constraints: Short-wave infrared (SWIR) is set to 1–3 μm, mid-wave infrared (MWIR) to 3–5 μm, and long-wave infrared (LWIR) to 8–12 μm. The bandgap threshold is determined by the highest photon energy at the short-wave end of each band. Determine that h is Planck's constant, c is the speed of light in vacuum, e is the elementary charge, and λ is the value of λ. min The shortwavelength of the target band is divided by e to convert Joules to electron volts; the resulting bandgap thresholds for the three bands are 1.240 eV, 0.413 eV, and 0.155 eV, respectively; the bandgap score (i.e., the bandgap constraint margin) The score obtained after linear soft thresholding normalization) is as follows: Calculation: When Eg = Ethreshold, the score is 0.5; below the threshold of 0.1 eV, the score is 0; above the threshold of 0.1 eV, the score is 1. Transparency score is calculated as follows: calculate; This means that x is restricted to the range of 0 to 1.

[0162] Conductivity evaluation: For the 6007 predicted log10(σ / τ), the 1st percentile (0.0693) and 99th percentile (2.2378) were used. Normalization. The materials were determined to be feasible materials, and calculations were performed. (i.e., the product of two 0-to-1 normalized scores); the IR-FOM of infeasible materials is set to 0.

[0163] Multi-objective optimization: Within the feasible region of each band, Pareto non-dominance judgment is performed with the goal of maximizing both transparency and conductivity evaluation quantities. Materials not dominated by other materials in both objectives are marked as Pareto fronts. Subsequently, within the Pareto fronts, materials are sorted twice from high to low according to IR-FOM, and the following are output: material_id (a unique material identifier assigned to each crystal structure record by the MaterialsProject database, e.g., mp-756006; different crystal structures with the same chemical formula can correspond to different material_ids), chemical formula, predicted Eg, predicted n, predicted log10(σ / τ), λp, three scores, feasible label, Pareto label, and IR-FOM.

[0164] Filtering results. For example... Figure 5As shown, the number of feasible materials for SWIR, MWIR, and LWIR are 2405, 1285, and 6, respectively, and the number of Pareto front materials are 10, 6, and 2, respectively. The top-ranked material in SWIR is mp-756006 (LiSbS), with predicted Eg=1.371 eV, n=4.467, log10(σ / τ)=1.852, λp=4.109 μm, and IR-FOM=0.715; the top-ranked material in MWIR is mp-1029469 (ZnSnN2), corresponding to 0.691 eV, 10.622, 1.843, 9.875 μm, and 0.818; and the top-ranked material in LWIR is mp-4964 (YSbPt), corresponding to 0.405 eV, 24.550, 1.761, 25.088 μm, and 0.780. All three records are located at the Pareto front of their respective bands.

[0165] Figure 3 This indicates that, under the data and parameters of this embodiment, XGBoost's prediction performance is close to that of Random Forest and better than that of the multilayer perceptron baseline; Figure 4 This indicates that the three independent models can learn the descriptor contributions related to the target performance, respectively. Figure 5 This indicates that as the target band expands from SWIR to LWIR, the number of materials that meet both transparency and conductivity thresholds decreases significantly, and Pareto sorting can preserve the different optimal trade-offs between transparency and conductivity.

[0166] The candidate materials in this embodiment are outputs of machine learning and physical constraint models and are not equivalent to having been verified through sample experiments. For candidate materials with high predicted refractive index values, strong absorption, strong damping, or significant anisotropy, subsequent verification should be performed using complex dielectric function, dielectric tensor, or first-principles optical calculations, and further validated through sample preparation, infrared transmittance, sheet resistance, and carrier transport tests.

Claims

1. A method for screening infrared transparent conductive materials by wavelength based on machine learning and multi-objective optimization, characterized in that... The aforementioned method for screening infrared transparent conductive materials by wavelength band is implemented according to the following steps: Step 1: Combine Magpie elemental property statistics, different stoichiometric norms, average s, p, d and f valence electron counts, material ionicity feasibility markers, maximum ionicity and average ionicity into a material descriptor. Use the material descriptor as input and the material band gap as the prediction target to train the band gap prediction model and obtain the trained band gap prediction model. Step 2: Using the material descriptor as input and the material refractive index as the prediction target, train the refractive index prediction model to obtain the trained refractive index prediction model. Step 3: Using the material descriptor as input and the material electronic transport parameter S as the prediction target, train the electronic transport parameter prediction model to obtain the trained electronic transport parameter prediction model. Step 4: Convert the material to be screened into the corresponding material descriptor and input it into the trained bandgap prediction model, the trained refractive index prediction model and the trained electron transport parameter prediction model respectively to obtain the predicted bandgap Eg, the predicted refractive index n and the predicted electron transport parameter S respectively. Step 5: Obtain the high-frequency relative permittivity ε from the predicted refractive index n. ∞ ; When carrier damping is neglected, the angular frequency of free carrier plasma reflection is expressed as: When carrier damping is considered, the angular frequency of free carrier plasma reflection is expressed as: Where ε0 is the vacuum permittivity and c is the speed of light in vacuum. The carrier damping rate; The plasma reflection edge wavelength is determined according to the following formula: in, The vacuum permittivity, It is the speed of light in a vacuum. Step 6: Set the target infrared band B, which is represented as follows: in, The shortest wavelength in the target band. The longest wavelength in the target band; Establish preliminary constraints on interband absorption and free carrier plasma reflection edges; A set of candidate materials was obtained through preliminary screening using interband absorption confinement and free carrier plasma reflection edge confinement. Step 7: Set target wavelength ranges for short-wave infrared, mid-wave infrared and long-wave infrared respectively, and calculate the transparency evaluation value and conductivity evaluation value of each candidate material in the candidate material set in the corresponding target infrared band B. The transparency evaluation parameters are bandgap constraint margin, plasma reflection edge constraint margin, or single-interface reflectivity. The bandgap constraint margin Represented as: ； The plasma reflection edge constraint margin Represented as: ； The single interface reflectivity Represented as: ； The conductivity evaluation quantity is the predicted electron transport parameter S, and n is the predicted refractive index; Step 8: Using transparency and conductivity as optimization objectives, perform Pareto non-dominated sorting on the candidate material set to obtain the Pareto frontier materials; Step 9: Calculate the infrared comprehensive evaluation parameter IR-FOM for Pareto front materials. B And according to the infrared comprehensive evaluation parameter IR-FOM B Secondary sorting and infrared comprehensive evaluation parameters IR-FOM B Represented as: in, For the target infrared band The corresponding transparency evaluation metric, For the target infrared band The corresponding conductivity evaluation quantity, An evaluation function that simultaneously considers both transparency and conductivity evaluation metrics; After secondary sorting, the selected infrared transparent conductive materials are obtained, thus completing the band-based screening method for infrared transparent conductive materials based on machine learning and multi-objective optimization.

2. The method for band-specific screening of infrared transparent conductive materials based on machine learning and multi-objective optimization according to claim 1, characterized in that... In step one, the bandgap prediction model is either an XGBoost regression model or a random forest model.

3. The method for band-specific screening of infrared transparent conductive materials based on machine learning and multi-objective optimization according to claim 1, characterized in that... In step two, the refractive index prediction model is either the XGBoost regression model or the random forest model.

4. The method for band-specific screening of infrared transparent conductive materials based on machine learning and multi-objective optimization according to claim 1, characterized in that... In step three, the prediction model for electron transport parameters is either an XGBoost regression model or a random forest model.

5. The method for band-specific screening of infrared transparent conductive materials based on machine learning and multi-objective optimization according to claim 1, characterized in that... When the XGBoost regression model is used for the bandgap prediction model, refractive index prediction model, and electron transport parameter prediction model, the number of regression trees is set to 400, and the maximum depth of a single tree is set to 5.

6. The method for band-specific screening of infrared transparent conductive materials based on machine learning and multi-objective optimization according to claim 5, characterized in that... The learning rate in the XGBoost regression model is set to 0.

05.

7. The method for band-specific screening of infrared transparent conductive materials based on machine learning and multi-objective optimization according to claim 1, characterized in that... In step three, the electron transport parameter S is the conductivity. Ratio to carrier relaxation time τ: 。 8. The method for band-specific screening of infrared transparent conductive materials based on machine learning and multi-objective optimization according to claim 1, characterized in that... In step five, the predicted refractive index and the high-frequency relative permittivity satisfy the following: Where, ε ∞ is the high-frequency relative permittivity, and n is the predicted refractive index.

9. The method for band-specific screening of infrared transparent conductive materials based on machine learning and multi-objective optimization according to claim 1, characterized in that... The expression for establishing the interband absorption constraint in step six is ​​as follows: Where h is Planck's constant and c is the speed of light in a vacuum.

10. The method for band-specific screening of infrared transparent conductive materials based on machine learning and multi-objective optimization according to claim 1, characterized in that... Step nine involves the infrared comprehensive evaluation parameter IR-FOM. B Represented as: 。

Images