Duplex stainless steel and duplex stainless steel design method based on machine learning

Abstract

The invention relates to the technical field of duplex stainless steel design, and provides duplex stainless steel and a duplex stainless steel design method based on machine learning, the design method comprises the following steps: step 1, obtaining data of the duplex stainless steel, forming a database, and dividing into a training set, a verification set and a test set; 2, carrying out feature engineering based on the database formed in the step 1, and calculating a chromium equivalent, a nickel equivalent and a pitting corrosion resistance equivalent; step 3, establishing a yield strength prediction model by using a machine learning model, and training by using the training set data; 4, evaluating the trained final yield strength prediction model by using the test set; and 5, taking the trained yield strength prediction model as a target function, and performing global optimization in a multi-dimensional component space by adopting a non-dominated sorting genetic algorithm to obtain duplex stainless steel candidate components. According to the method, the accuracy and efficiency of component design of the high-toughness duplex stainless steel can be improved.

Description

Technical Field

[0001] This invention relates to the field of duplex stainless steel design technology, and more particularly to a duplex stainless steel and a machine learning-based duplex stainless steel design method. Background Technology

[0002] Duplex stainless steel (DSS) combines the excellent toughness and weldability of austenitic stainless steel with the high strength and resistance to chloride stress corrosion cracking of ferritic stainless steel due to its unique duplex microstructure—approximately equal proportions of austenite (γ) and ferrite (α) phases. This characteristic makes it a key structural material in harsh corrosive environments such as oil and gas, chemical industry, marine engineering, nuclear power, and desalination environmental protection, possessing enormous economic value and strategic significance.

[0003] Traditional compositional design of duplex stainless steel relies heavily on a trial-and-error approach. This process, typically based on limited physical metallurgy knowledge and researchers' experience, involves a lengthy cycle of design, smelting, rolling, heat treatment, performance testing, and microstructural observation to gradually approach the target. This method is time-consuming, generates a large number of scraps, leads to high experimental costs, and has low efficiency. Therefore, exploring the vast compositional design space of duplex stainless steel efficiently, accurately, and at low cost, and developing high-strength and high-toughness duplex stainless steel, has significant theoretical and practical value. Summary of the Invention

[0004] This invention addresses the technical problems of traditional duplex stainless steel composition design, which is often based on limited physical metallurgy knowledge and researchers' experience, resulting in long development cycles, a large number of scraps, high testing costs, and low testing efficiency. It proposes a duplex stainless steel and a machine learning-based duplex stainless steel design method, aiming to shorten the development cycle, reduce development costs, and improve the accuracy and efficiency of high-strength and high-toughness duplex stainless steel composition design.

[0005] This invention provides a duplex stainless steel, wherein the duplex stainless steel material is Fe a Cr b Ni c Mo d Mn e Si f N g W h C i ;

[0006] Where a is the mass percentage of Fe, b is the mass percentage of Cr, c is the mass percentage of Ni, d is the mass percentage of Mo, e is the mass percentage of Mn, f is the mass percentage of Si, g is the mass percentage of N, h is the mass percentage of W, and i is the mass percentage of C; b = 18~25wt%, c = 2~8wt%, d = 2.5~4.0wt%, e = 5~12wt%, f = 0.2~0.8%, g = 0.1~0.5wt%, h = 0~2wt%, i = 0.01~0.03wt%, and the balance is a.

[0007] Preferably, the duplex stainless steel has a yield strength of 412 MPa to 886 MPa, a tensile strength of 543 MPa to 1062 MPa, and an elongation of 13.2% to 25.1%.

[0008] Correspondingly, the present invention provides a machine learning-based design method for duplex stainless steel according to any embodiment of the present invention, comprising the following processes:

[0009] Step 1: Obtain data on duplex stainless steel, form a database, and divide it into training, validation, and test sets; the data characteristics of the duplex stainless steel include composition, preparation process, and mechanical properties;

[0010] Step 2: Based on the database formed in Step 1, perform feature engineering to calculate the chromium equivalent, nickel equivalent, and pitting corrosion resistance equivalent.

[0011] Creq = %Cr + %Mo + 0.7×%Nb + 1.5×%Si;

[0012] Nieq = %Ni + 35×%C + 20×%N + 0.25×%Cu;

[0013] PREN = %Cr + 3.3×%Mo + 16×%N;

[0014] Wherein, Creq represents chromium equivalent, Nieq represents nickel equivalent, PREN represents pitting resistance equivalent, %Cr represents the mass percentage of Cr, %Mo represents the mass percentage of Mo, %Nb represents the mass percentage of Nb, %Si represents the mass percentage of Si, %Ni represents the mass percentage of Ni, %C represents the mass percentage of C, %N represents the mass percentage of N, and %Cu represents the mass percentage of Cu.

[0015] Step 3: Utilize machine learning models to establish yield strength prediction models and train them using training set data;

[0016] Step 4: Evaluate the trained final yield strength prediction model using the test set;

[0017] Step 5: Using the trained yield strength prediction model as the objective function, a non-dominated sorting genetic algorithm is used for global optimization in the multidimensional component space to obtain candidate components for duplex stainless steel.

[0018] Furthermore, the machine learning model includes Gradient Boosting Decision Tree (XGBoost), Random Forest (RF), Support Vector Machine (SVM), or artificial neural network.

[0019] Furthermore, the yield strength prediction model takes elemental characteristics, empirical parameter characteristics and mechanical properties, chromium equivalent, nickel equivalent and pitting corrosion resistance equivalent as inputs, and the yield strength value as the output target.

[0020] The empirical parameter characteristics of the yield strength prediction model include: valence electron concentration (VEC), electronegativity (Δχ), atomic size difference (δ), and enthalpy of mixing (ΔH). mix Mixed entropy ΔS mix Atomic stacking parameter Λ, free electron concentration e / a, electron work function w, cohesive energy Ec, radius R, critical shear stress F, lattice distortion energy U, van der Waals radius A, Pauling electronegativity B, modulus mismatch factor η, shear modulus difference ΔG, shear modulus G, first ionization energy E;

[0021] The identified elemental characteristics include: Fe, Cr, Ni, Mo, Mn, Si, N, W, and C.

[0022] Furthermore, in step 5, the optimization objective is set to maximize the yield strength, and the constraint condition is set to YS≥400MPa, where YS is the yield strength. By controlling the Creq / Nieq ratio, the calculated phase ratio is ensured to be between 40% and 60%. The composition range constraints are set as follows: Cr: 18-25%, Ni: 2-8%, Mo: 2.5-4.0%, N: 0.1-0.5%, C≤0.03%, Mn: 5-12%, Si: 0.2-0.8%. The trained yield strength prediction model is used as the objective function, and a non-dominated sorting genetic algorithm is used for global optimization in the multidimensional composition space.

[0023] Correspondingly, the present invention provides a method for preparing duplex stainless steel based on a machine learning-based duplex stainless steel design method according to any embodiment of the present invention, comprising the following steps:

[0024] From the candidate components obtained by the machine learning-based duplex stainless steel design method provided in any embodiment of the present invention, the target component is selected, and iron blocks, chromium blocks, nickel blocks, molybdenum blocks, manganese blocks, silicon blocks and carbon powder are weighed out. The nitrogen element is added in the form of CrN. The proportion of each element in the alloy is a mass percentage.

[0025] The weighed raw materials are evenly mixed and placed in a water-cooled crucible of a vacuum melting furnace. After the vacuum melting furnace is evacuated, it is filled with inert gas for protection. The induced current is adjusted to completely melt the elements. After maintaining the liquid state for a period of time, it is cooled to obtain duplex stainless steel with the target composition.

[0026] Furthermore, the iron blocks, chromium blocks, nickel blocks, molybdenum blocks, manganese blocks, tungsten blocks, silicon blocks, carbon powder, and chromium nitride are all industrial-grade pure raw materials with a purity of 99.5 wt.% or higher.

[0027] Furthermore, during the vacuum arc melting process, the vacuum is evacuated to 3×10⁻⁶. -3 Pa ~ 5.5 × 10 -3 Pa, then backflushing argon gas to about 0.05~0.06 Pa; the vacuum arc melting temperature is 2100℃ ~ 2300℃; the voltage for each arc ignition melting is 10-15V, and the current for each arc ignition melting is 300-350A.

[0028] Furthermore, during vacuum arc melting, the ingot is flipped and melted six to eight times, with each melting session lasting one to two minutes.

[0029] This invention provides a duplex stainless steel and a machine learning-based design method for duplex stainless steel. This invention also provides an Fe... a Cr b Ni c Mo d Mn e Si f N g W h C iDuplex stainless steel materials, including this series, all contain ferrite and austenite in their structure, exhibiting excellent mechanical properties. This invention presents a machine learning-based design method for duplex stainless steel. First, a high-performance duplex stainless steel composition-performance database is constructed, along with feature engineering and datasets. Then, a machine learning model is established and trained. Next, the model's performance is verified and evaluated, and the algorithm is optimized to obtain the target composition. This method achieves intelligent and targeted design of duplex stainless steel compositions, significantly shortening the R&D cycle, reducing R&D costs, and efficiently obtaining novel compositions with superior performance that are difficult to discover using traditional methods. It significantly improves the accuracy and efficiency of high-strength and high-toughness duplex stainless steel composition design. The prepared duplex stainless steels are all ferrite and austenitic, exhibiting high strength and high plasticity mechanical properties, as well as excellent corrosion resistance. This solves the problem of developing high-strength and high-toughness duplex stainless steels to meet more demanding service environments. This invention overcomes the drawbacks of traditional material development, which heavily relies on expert experience and numerous repetitive experiments. By establishing a quantitative mapping relationship between composition, process, and performance through a machine learning model, candidate compositions can be accurately derived from the target performance, greatly reducing the blindness of experiments. Attached Figure Description

[0030] Figure 1 This is the overall flowchart of the machine learning-based duplex stainless steel design method in Example 2;

[0031] Figure 2 A schematic diagram of the intensity prediction results provided in Example 2;

[0032] Figure 3 The tensile engineering stress-engineering strain curve of the high-strength and tough duplex stainless steel prepared in Example 3 is shown. Detailed Implementation

[0033] To make the technical problems solved by this invention, the technical solutions adopted, and the technical effects achieved clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings, not all of them.

[0034] Example 1

[0035] This invention provides a duplex stainless steel, wherein the duplex stainless steel material is Fe a Cr b Ni c Mo d Mn e Si f N g W h Ci ;

[0036] Where a is the mass percentage of Fe, b is the mass percentage of Cr, c is the mass percentage of Ni, d is the mass percentage of Mo, e is the mass percentage of Mn, f is the mass percentage of Si, g is the mass percentage of N, h is the mass percentage of W, and i is the mass percentage of C; b=23wt%, c=6wt%, d=3wt%, e=6wt%, f=0.5%, g=0.4wt%, h=0.5wt%, i=0.02wt%, and the balance is a.

[0037] The duplex stainless steel has a yield strength of 656 MPa, a tensile strength of 841 MPa, and an elongation of 19.3%.

[0038] Example 2

[0039] This invention provides a machine learning-based design method for duplex stainless steel according to any embodiment of the invention, comprising the following steps:

[0040] Step 1: Obtain data on duplex stainless steel, form a database, and divide it into training, validation, and test sets.

[0041] Data characteristics for duplex stainless steel include composition (wt%), manufacturing process, and mechanical properties. Specifically, complete data entries for 283 commercial and experimental duplex stainless steels were collected from over 150 published documents and patents. The manufacturing process for the collected duplex stainless steels was vacuum arc melting, and the mechanical property testing method was room temperature uniaxial tensile testing.

[0042] The collected data was cleaned, and records with missing key data or obvious anomalies were removed, resulting in 256 valid data samples, which formed the database. The 256 samples were randomly divided into a training set (205 samples), a validation set (25 samples), and a test set (26 samples).

[0043] Step 2: Based on the database formed in Step 1, perform feature engineering to calculate the chromium equivalent, nickel equivalent, and pitting corrosion resistance equivalent.

[0044] Creq = %Cr + %Mo + 0.7×%Nb + 1.5×%Si;

[0045] Nieq = %Ni + 35×%C + 20×%N + 0.25×%Cu;

[0046] PREN = %Cr + 3.3×%Mo + 16×%N;

[0047] Wherein, Creq represents chromium equivalent, Nieq represents nickel equivalent, PREN represents pitting resistance equivalent, %Cr represents the mass percentage of Cr, %Mo represents the mass percentage of Mo, %Nb represents the mass percentage of Nb, %Si represents the mass percentage of Si, %Ni represents the mass percentage of Ni, %C represents the mass percentage of C, %N represents the mass percentage of N, and %Cu represents the mass percentage of Cu.

[0048] Step 3: Utilize a machine learning model to establish a yield strength prediction model and train it using the training set data.

[0049] The machine learning models include XGBoost (eXtreme Gradient Boosting), Random Forest (RF), Support Vector Machine (SVM), or Artificial Neural Network (ANN).

[0050] The yield strength (YS) prediction model takes elemental characteristics, empirical parameter characteristics, mechanical properties, chromium equivalent, nickel equivalent, and pitting resistance equivalent as inputs, and the yield strength (YS) value as the output target.

[0051] The empirical parameter features of the yield strength (YS) prediction model include: valence electron concentration VEC, electronegativity Δχ, atomic size difference δ, and enthalpy of mixing ΔH. mix Mixed entropy ΔS mix Atomic stacking parameter Λ, free electron concentration e / a, electron work function w, cohesive energy Ec, radius R, critical shear stress F, lattice distortion energy U, van der Waals radius A, Pauling electronegativity B, modulus mismatch factor η, shear modulus difference ΔG, shear modulus G, first ionization energy E;

[0052] The identified elemental characteristics include: Fe, Cr, Ni, Mo, Mn, Si, N, W, and C.

[0053] After training the yield strength prediction model, the maximum depth, learning rate, and subsampling ratio of the model are tuned using Bayesian optimization on a validation set to minimize the root mean square error (RMSE) of the predictions.

[0054] Step 4: Evaluate the trained final yield strength prediction model using the test set.

[0055] The coefficient of determination R² for the yield strength (YS) prediction model on the test set is 0.93, calculated as R² = 1 - (SSR / SST), where SST is the total sum of squares, representing the volatility of the actual data, and SSR is the residual sum of squares, representing the model's prediction error. Figure 2 As shown, the results indicate that the model has extremely high prediction accuracy.

[0056] Step 5: Using the trained yield strength (YS) prediction model as the objective function, a non-dominated sorting genetic algorithm is used to perform global optimization in the multidimensional component space to obtain candidate components for duplex stainless steel.

[0057] Specifically, the optimization objective is to maximize the yield strength (YS), with the constraint that YS ≥ 400 MPa. YS represents the yield strength. By controlling the Creq / Nieq ratio, the calculated proportion is ensured to be between 40% and 60%. The composition range constraints are set as follows: Cr: 18-25%, Ni: 2-8%, Mo: 2.5-4.0%, N: 0.1-0.5%, C≤0.03%, Mn: 5-12%, Si: 0.2-0.8% (wt%). The trained yield strength (YS) prediction model is used as the objective function, and a non-dominated sorting genetic algorithm is employed for global optimization in the multidimensional composition space. The population size is set to 100, and the iterations are performed for 200 generations.

[0058] From the candidate components obtained by the optimization algorithm, the better component can be selected as the target component for actual preparation. Experimental data obtained from the smelting experiment can be fed back into the database to update the database.

[0059] Example 3

[0060] This invention provides a method for preparing duplex stainless steel based on a machine learning-based duplex stainless steel design method according to any embodiment of the invention, comprising the following steps:

[0061] From the candidate compositions obtained by the machine learning-based duplex stainless steel design method provided in any embodiment of the present invention, the target composition Fe is selected. 55.93 Cr 25 Ni4Mo 2.5 Mn 12 Si 0.4 N 0.15 C 0.02 Iron, chromium, nickel, molybdenum, manganese, silicon, and carbon powder are weighed out, with nitrogen added in the form of CrN. The proportions of each element in the alloy are by mass percentage. The iron, chromium, nickel, molybdenum, manganese, tungsten, silicon, and chromium nitride are all industrial-grade pure raw materials with a purity of 99.5 wt.% or higher.

[0062] The weighed raw materials are evenly mixed and placed in a water-cooled crucible of a vacuum melting furnace. After the vacuum melting furnace is evacuated, it is filled with inert gas for protection. The induced current is adjusted to completely melt the elements. After maintaining the liquid state for a period of time, it is cooled to obtain duplex stainless steel with the target composition.

[0063] During vacuum arc melting, the vacuum is evacuated to 3×10⁻⁶. -3 Pa ~ 5.5 × 10 -3 Pa, then backflush with argon gas to approximately 0.05 MPa. Evacuate to a vacuum level of 3 × 10⁻⁶ MPa. -3 When the pressure reaches 0.05 Pa, the argon gas is backflushed to 0.05 Pa, and the smelting process begins.

[0064] The vacuum arc melting temperature is 2100℃ ~ 2300℃.

[0065] The voltage for each arc ignition melting is 10-15V, and the current for each arc ignition melting is 300-350A.

[0066] Before smelting duplex stainless steel, Ti alloy ingots are smelted first to absorb residual oxygen in the vacuum arc furnace.

[0067] During vacuum arc melting, the ingot is turned over and melted six to eight times, each time for one to two minutes, to ensure uniform composition. After cooling, duplex stainless steel ingots are obtained.

[0068] The mechanical properties of the duplex stainless steel tensile sheets prepared in this embodiment were tested. The results showed a yield strength of 825 MPa, a breaking strength of 986 MPa, and an elongation of 25%. The tensile stress-strain curve of the duplex stainless steel ingot under as-cast conditions is shown below. Figure 3 As shown.

[0069] Example 4

[0070] This invention provides a method for preparing duplex stainless steel based on the machine learning-based duplex stainless steel design method provided in any embodiment of this invention. The preparation process is consistent with that in Example 3. In this embodiment, the target component Fe is selected from the candidate components obtained by the machine learning-based duplex stainless steel design method provided in any embodiment of this invention. 60.18 Cr 23 Ni5Mo 2.5 Mn8Si 0.5 N 0.3 W 0.5 C 0.02 .

[0071] The mechanical properties of the duplex stainless steel cut tensile sheet prepared in this embodiment were tested. The test results showed that its yield strength was 694 MPa, its breaking strength was 872 MPa, and its elongation was 20.4%.

[0072] 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 to the technical solutions described in the foregoing embodiments, or equivalent substitutions for some or all of the technical features, 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 duplex stainless steel, characterized in that, The duplex stainless steel material is Fe. a Cr b Ni c Mo d Mn e Si f N g W h C i ; Where a is the mass percentage of Fe, b is the mass percentage of Cr, c is the mass percentage of Ni, d is the mass percentage of Mo, e is the mass percentage of Mn, f is the mass percentage of Si, g is the mass percentage of N, h is the mass percentage of W, and i is the mass percentage of C; b = 18~25wt%, c = 2~8wt%, d = 2.5~4.0wt%, e = 5~12wt%, f = 0.2~0.8%, g = 0.1~0.5wt%, h = 0~2wt%, i = 0.01~0.03wt%, and the balance is a.

2. The duplex stainless steel according to claim 1, characterized in that, The duplex stainless steel has a yield strength of 412 MPa to 886 MPa, a tensile strength of 543 MPa to 1062 MPa, and an elongation of 13.2% to 25.1%.

3. A machine learning-based design method for duplex stainless steel according to claim 2, characterized in that, Includes the following processes: Step 1: Obtain data on duplex stainless steel, form a database, and divide it into training, validation, and test sets; the data characteristics of the duplex stainless steel include composition, preparation process, and mechanical properties; Step 2: Based on the database formed in Step 1, perform feature engineering to calculate the chromium equivalent, nickel equivalent, and pitting corrosion resistance equivalent. Creq = %Cr + %Mo + 0.7×%Nb + 1.5×%Si; Nieq = %Ni + 35×%C + 20×%N + 0.25×%Cu; PREN = %Cr + 3.3×%Mo + 16×%N; Wherein, Creq represents chromium equivalent, Nieq represents nickel equivalent, PREN represents pitting resistance equivalent, %Cr represents the mass percentage of Cr, %Mo represents the mass percentage of Mo, %Nb represents the mass percentage of Nb, %Si represents the mass percentage of Si, %Ni represents the mass percentage of Ni, %C represents the mass percentage of C, %N represents the mass percentage of N, and %Cu represents the mass percentage of Cu. Step 3: Utilize machine learning models to establish yield strength prediction models and train them using training set data; Step 4: Evaluate the trained final yield strength prediction model using the test set; Step 5: Using the trained yield strength prediction model as the objective function, a non-dominated sorting genetic algorithm is used for global optimization in the multidimensional component space to obtain candidate components for duplex stainless steel.

4. The machine learning-based design method for duplex stainless steel according to claim 3, characterized in that, The machine learning models include Gradient Boosting Decision Tree (XGBoost), Random Forest (RF), Support Vector Machine (SVM), or artificial neural networks.

5. The machine learning-based duplex stainless steel design method according to claim 3, characterized in that, The yield strength prediction model takes elemental characteristics, empirical parameter characteristics, mechanical properties, chromium equivalent, nickel equivalent, and pitting corrosion resistance equivalent as inputs, and the yield strength value as the output target. The empirical parameter characteristics of the yield strength prediction model include: valence electron concentration (VEC), electronegativity (Δχ), atomic size difference (δ), and enthalpy of mixing (ΔH). mix Mixed entropy ΔS mix Atomic stacking parameter Λ, free electron concentration e / a, electron work function w, cohesive energy Ec, radius R, critical shear stress F, lattice distortion energy U, van der Waals radius A, Pauling electronegativity B, modulus mismatch factor η, shear modulus difference ΔG, shear modulus G, first ionization energy E; The identified elemental characteristics include: Fe, Cr, Ni, Mo, Mn, Si, N, W, and C.

6. The machine learning-based design method for duplex stainless steel according to claim 3, characterized in that, In step 5, the optimization objective is set to maximize the yield strength, and the constraint condition is YS≥400MPa, where YS is the yield strength. By controlling the Creq / Nieq ratio, the calculated phase ratio is ensured to be between 40% and 60%. The composition range constraints are set as follows: Cr: 18-25%, Ni: 2-8%, Mo: 2.5-4.0%, N: 0.1-0.5%, C≤0.03%, Mn: 5-12%, Si: 0.2-0.8%. The trained yield strength prediction model is used as the objective function, and a non-dominated sorting genetic algorithm is used for global optimization in the multidimensional composition space.

7. A method for preparing duplex stainless steel based on a machine learning-based duplex stainless steel design method according to any one of claims 3-7, characterized in that, Includes the following processes: From the candidate components obtained by the machine learning-based duplex stainless steel design method according to any one of claims 3-7, select the target component, weigh iron blocks, chromium blocks, nickel blocks, molybdenum blocks, manganese blocks, silicon blocks and carbon powder, and add nitrogen element in the form of CrN, wherein the proportion of each element in the alloy is a mass percentage. The weighed raw materials are evenly mixed and placed in a water-cooled crucible of a vacuum melting furnace. After the vacuum melting furnace is evacuated, it is filled with inert gas for protection. The induced current is adjusted to completely melt the elements. After maintaining the liquid state for a period of time, it is cooled to obtain duplex stainless steel with the target composition.

8. The method for preparing duplex stainless steel according to claim 7, characterized in that, Iron blocks, chromium blocks, nickel blocks, molybdenum blocks, manganese blocks, tungsten blocks, silicon blocks, carbon powder, and chromium nitride are all industrial-grade pure raw materials with a purity of 99.5 wt.% or higher.

9. The method for preparing duplex stainless steel according to claim 7, characterized in that, During vacuum arc melting, the vacuum is evacuated to 3×10⁻⁶. -3 Pa ~ 5.5 × 10 -3 Pa, then backflushing argon gas to about 0.05~0.06 Pa; the vacuum arc melting temperature is 2100℃ ~ 2300℃; the voltage for each arc ignition melting is 10-15V, and the current for each arc ignition melting is 300-350A.

10. The method for preparing duplex stainless steel according to claim 7, characterized in that, During vacuum arc melting, the ingot is turned over and melted six to eight times, with each melting lasting one to two minutes.

Images