Aluminum alloy melt pouring simulation method and system

Abstract

The application discloses an aluminum alloy melt pouring simulation method and system, comprising: solving the basic physical field of pouring filling and solidification process; discretizing the free surface into surface material entities corresponding to one film piece respectively, taking the surface material entities as nodes and the adjacent film piece communication relationship as edges to construct a dynamic surface material entity graph, and creating a new entity at a newly generated free surface, so that the individual identity and folding topology of the film piece are continued across time steps; inputting the graph into a topology event identification network of a graph neural network, outputting event types and film piece pairs of self-folding adhesion, liquid front merging and injection flow fracture; inputting the film piece pairs and contact geometry into a configuration state reconstruction network, outputting the double-layer nature, welding state, orientation and maturity of the involved film pieces; Lagrangian advancing the film entity with the configuration state and freezing when the solid phase rate reaches a threshold; counting and generating a harmful double-layer film distribution field and combining a stress field to predict casting defects, and outputting orientation-resolved defect prediction basis in each part of the casting according to the defects.

Description

Technical Field

[0001] This application relates to the field of aluminum alloy casting process simulation technology, specifically to an aluminum alloy melt pouring simulation method and system, and particularly to numerical simulation of the aluminum alloy melt pouring filling and solidification process, casting defect prediction and process parameter optimization. Background Technology

[0002] Aluminum alloy casting is the fundamental forming method for obtaining complex aluminum alloy components. The mechanical properties and reliability of castings largely depend on the types, quantities, and distribution of casting defects formed during the pouring, filling, and solidification processes. To predict defect distribution before actual production and optimize process parameters such as pouring temperature, holding time, gate settings, and filling speed accordingly, numerical simulation is commonly used in this field to reproduce the pouring, filling, and solidification processes of molten aluminum alloy. Existing pouring simulations typically use the finite volume method or finite element method to solve the continuity, momentum, and energy equations of the melt, and use the fluid volume method or level set method to track the free surface of the melt, thereby outputting bulk physical fields such as the filling velocity field, pressure field, temperature field, and solidity field. Empirical or semi-empirical criteria are then added to these physical fields to predict defects such as shrinkage porosity, shrinkage cavity, hot cracking, and cold shuts. In recent years, machine learning methods have also been used to accelerate the solution of the above-mentioned bulk physical fields or to directly predict shrinkage porosity and shrinkage cavity behavior and scalarize defect indices.

[0003] Aluminum alloy molten metal is highly chemically reactive. Once its free surface is exposed to an oxygen-containing atmosphere, an oxide film instantly forms, continuously regenerating as the free surface expands and as the melt flows. During the casting process, the molten metal enters through the gate and propagates along the runner and mold cavity. Its free surface is in a state of intense turbulence, sometimes bulging, splashing, and churning, and sometimes colliding and merging with two other liquid streams at the point where it falls back to the surface or where multiple streams meet. In these moments of folding and merging of the free surface, the oxide film that originally covered the liquid surface is incorporated into the interior of the melt. The severity of the damage caused by an entrained oxide film to a casting depends not on the quantity or location of the entrained film, but on its specific configuration at the time of entrainment. Only when two oxide films adhere to each other with their non-wetting surfaces, enclosing an air gap in between, and failing to re-weld with the surrounding melt, do they form a folded double-layer film configuration, which then becomes a source of crack and porosity after solidification. A single-layer, flat entrained film that can be wetted by the melt is relatively harmless. Furthermore, the spatial orientation of the entrained film relative to the subsequent principal stress direction of the casting, as well as the exposure time experienced by the film from its formation to its entrainment, jointly determine the thickness, strength, and maturity of the entrained gas components of the oxide film, and thus jointly determine whether the defect eventually develops into a harmful defect and the severity of its damage.

[0004] To predict oxide film entrapment, existing casting simulation methods typically rely on fluid volume analysis to track the free surface, then use kinematic approximations combined with Boolean logic to determine whether melt turbulence is sufficient to trigger oxide film entrapment. Alternatively, they may characterize the amount of entrapped oxide with scalars flowing with the melt and correlate the exposure time of the free surface with the flow path length to qualitatively predict the final location of the entrapped oxide. Other approaches use discrete particles instead of oxide inclusions and approximate the spatial distribution of entrapped oxide by tracking particle position distribution. However, these methods of characterizing entrapped oxide with scalar fields or Boolean decisions treat the entrapped oxide as grid-by-grid scalars or isolated tracer particles, erasing the individual identity and folded topological relationships of the film as a continuous material surface. They cannot answer whether an entrapped oxide film at a particular location is in a harmful double-layer unbonded folded state, nor can they provide the spatial orientation of the film relative to the casting coordinate system. Currently, for the criteria of cold shut-off and confluence at the junction of liquid fronts, existing technologies usually only mark the position or time when the liquid surfaces meet, without tracking whether the oxide films of the two surfaces are welded together or the joint configuration. Therefore, in the case of multiple liquid fronts converging to form a joint, existing technologies have difficulty in determining whether an unwelded and harmful configuration has been formed at the joint.

[0005] In recent years, machine learning methods such as graph neural networks have been introduced to accelerate the simulation of Lagrange free surface flows. Using particles or mesh nodes as graph nodes and interactions as edges, they are used to predict the velocity and pressure fields of fluids and handle interface topological changes under large deformations. However, the objects predicted by these graph neural network methods are still limited to physical quantities such as velocity and pressure of the bulk fluid. Their graph nodes represent the fluid particles themselves, rather than oxide films carrying the generation time and folding history. Therefore, they do not recognize the film pairing and the resulting bilayer configuration that occurs during free surface folding or liquid front merging—material surface events. Existing machine learning methods have not yet addressed the issue of oxide films as two-dimensional material entities with individual histories, and the evolution of their identity and folding topology during the filling process.

[0006] The root of the problem lies in the fact that the melt in the casting simulation is treated as a continuous bulk medium, and its free surface is merely a geometric interface at a certain moment, lacking a piecewise identity that can continue across time. Therefore, when the free surface folds over and adheres to itself, or when two propelling liquid fronts collide and merge, the record of which part of the surface becomes an unbonded double film, its orientation, and its maturity is immediately erased. Consequently, the casting simulation cannot output the configuration of the oxide film entrained throughout the solidified casting, and therefore cannot translate the quantity or area distribution of the entrained oxides into the actual distribution of the harmful double film. Furthermore, it is difficult to combine this distribution with the stress field inside the casting to determine the tendency for fatigue crack initiation. Existing materials science research has shown that bilayer films, as folded oxide films, enclose air gaps and lack interlayer bonding, constitute crack initiation sites, and their spatial orientation becomes the dominant factor in fracture as size increases. However, these understandings are all based on post-experimental observations and tests. Current casting simulation methods cannot directly calculate and output the distribution of harmful bilayer films with orientation-resolved and maturity-resolved characteristics during the filling process simulation.

[0007] The aforementioned problems are particularly acute in several special cases. One is in turbulent or unstable pouring conditions, such as when the pouring flow itself breaks or splashes, resulting in extremely frequent self-contact events on the free surface. Another is in cases where multiple liquid fronts converge within the mold cavity during the filling of thin-walled or multi-gate parts, leading to frequent collisions and merging of two liquid surfaces each carrying an oxide film, forming unwelded oxide seams. It is precisely in these situations, dominated by self-contact on the free surface and liquid front merging, that the loss of film configuration information is most severe. Existing pouring simulations are most incapable of providing results usable for defect identification in these cases, making it difficult to predict the failure behavior of castings poured under fatigue loads under such process conditions, and lacking reliable basis for optimizing process parameters. Therefore, how to maintain and restore the individual identity, fold topology, and maturity of the oxide film covering the free surface of the melt in pouring simulations, thereby accurately outputting the configuration state of the oxide film entrained throughout the solidified casting, is a technical problem urgently needing to be solved in this field. Summary of the Invention

[0008] To address the technical problem that existing technologies cannot reconstruct the oxide film configuration state of various parts of the casting interior using filling dynamics, this invention provides a method and system for simulating aluminum alloy melt pouring. The core of this invention lies in explicitly representing the oxide film covering the free surface of the melt as a family of surface material entities carrying individual identities and evolving during the filling process. A neural network is used to identify topological events such as surface folding and liquid front merging, and to reconstruct the resulting film configuration state. This allows for the output of the bilayer nature, spatial orientation, and maturity distribution of harmful double-layer films throughout the solidified casting interior, used for casting defect prediction and process parameter optimization.

[0009] One aspect of the present invention provides a method for simulating aluminum alloy melt casting, comprising: Step 1, solving the fundamental physical fields of the casting, filling, and solidification process, and outputting the geometric position of the free surface, the velocity and temperature of the melt at the free surface, and the velocity field and solidity field of the bulk melt step by step; Step 2, discretizing the free surface into surface material entities corresponding to a diaphragm, constructing a dynamic surface material entity graph with each surface material entity as a node and the connectivity between adjacent diaphragms as edges, wherein the nodes carry the position of the diaphragm, the unit normal vector, the cumulative exposure time since exposure to the atmosphere, and features identifying the wetting state of the two diaphragm surfaces, and creating new surface material entities at newly generated free surfaces, so that the individual identity and folding topology of the diaphragms continue across time steps; Step 3, inputting the dynamic surface material entity graph into a graph neural network. The network's topology event recognition network outputs the topology event types and participating diaphragm pairs via message passing, including free surface self-folding and bonding, liquid front collision and merging, and injection fracture and splashing. Step four involves reconstructing the network with the diaphragm pairs participating in each topology event and their contact geometry input configuration states, outputting the configuration state of the formed entangled diaphragm (whether it is a double-layer or single-layer film, whether the two surfaces of the double-layer film are welded, its spatial orientation relative to the casting coordinate system, and its maturity. Step five involves treating the entangled diaphragm with the configuration state as an internal film entity within the bulk phase, allowing it to undergo Lagrangian propagation with the velocity field of the bulk melt, freezing it along with the configuration state when the solidity at its location reaches a threshold. Step six involves statistically analyzing the configuration states of the film entities frozen at various locations in the casting, generating a harmful double-layer film distribution field, and combining it with the stress field to predict casting defects.

[0010] Specifically, in step one, based on the given mold geometry, gate location and process parameters, the continuity equation, momentum equation and energy equation of the melt are solved by the finite volume method or finite element method, and the free surface of the melt is tracked by the fluid volume method or horizontal set method. The geometric position of the free surface, the melt velocity and temperature at the free surface, and the velocity field and solid fraction field of the bulk melt are output step by step, providing a dynamic basis for the subsequent construction of the surface material entity map and the tracking film entity.

[0011] Furthermore, while solving the fundamental physical field, a surface age transport equation is also solved on the free surface, so that each point on the free surface carries the accumulated exposure time since its formation upon exposure to the atmosphere. The source term of the surface age transport equation determines the film thickness growth rate according to the local temperature and oxygen partial pressure of the free surface based on oxidation growth kinetics, making the exposure time correspond to the thickness and strength of the oxide film. Thus, the exposure time of newly exposed areas on the free surface accumulates from 0, while the exposure time of areas with long-term exposure is larger, corresponding to thicker, stronger, and more difficult-to-re-weld oxide films by the melt, providing physical quantities for subsequent determination of the weldability and maturity of the film.

[0012] Specifically, in step two, the free surface at the current time step is discretized into a set of surface material entities. Each surface material entity corresponds to a membrane on the free surface, serving as a node in the dynamic surface material entity graph. The feature vector carried by the node includes at least the spatial position, area, unit normal vector, exposure time, local temperature, and the wetting state indicating that its two membrane surfaces face inwards towards the melt and outwards towards the air gap, respectively. Adjacent membranes are connected by edges, which characterize the connectivity of the membranes on the free surface and carry the relative orientation between the membranes. As the free surface expands due to the melt propagation, new surface material entities are created at the newly generated free surface, and their generation time is recorded at the current time. Simultaneously, their connectivity edges with existing membranes are updated, allowing the individual identity, generation time, and folding topology of each membrane to continue across time steps. Furthermore, each surface material entity undergoes Lagrangian propagation with the melt velocity at the free surface between time steps, and its area is updated according to the local stretching rate of the free surface, thereby superimposing the Lagrangian description of the free surface on top of the Eulerian description of the bulk melt.

[0013] Specifically, in step three, the dynamic surface material entity map is input into a topological event recognition network, which is a graph neural network. Through multiple rounds of message passing on the surface material entity map, the geometric and kinematic features of each diaphragm and its neighborhood are aggregated. The network outputs the topological events occurring or about to occur on the free surface and their types. These types include at least free surface self-folding and bonding, collision and merging of two propulsion fluid fronts, and injection flow fracture and splashing. The network also outputs the diaphragm pairs participating in each event and their contact geometry. The geometric and kinematic features include at least the distance between two diaphragms approaching each other, the angle between their unit normals, their relative approach velocity, and their respective exposure durations. Preferably, the topology event recognition network is trained in a supervised manner on a sample set generated by solving the basic physics field and tracking the high-resolution interface, and the topology event types and membrane pairing relationships are labeled. Further, for membrane pairs that are close to each other, when their unit normal vectors are opposite in direction and the distance is less than a set threshold, they are determined to be candidate events of self-folding and bonding or liquid front collision and merging, and are handed over to the topology event recognition network for further classification, so as to improve the detection rate of events in the case of dense self-contact events such as turbulent casting and multiple liquid front convergence.

[0014] Specifically, in step four, for each topological event identified in step three, a reconstruction network is constructed of the diaphragm pair participating in the event and its contact geometry input configuration state. This network outputs the configuration state of the wound diaphragm formed by the event. Specifically, the configuration state reconstruction network determines whether the formed diaphragm is a double-layer unwelded gas-filled folded state based on whether the membrane surfaces of the two diaphragms involved in the topological event are both non-wetting membrane surfaces, whether an air gap is sandwiched in the middle during the bonding, and the strength and weldability of the membrane determined by the exposure time of each of the two diaphragms; it determines the spatial orientation of the formed diaphragm relative to the casting coordinate system based on the geometric orientation of the bonding surfaces of the two diaphragms at the time of the event; it determines the maturity of the formed diaphragm based on the larger of the exposure time of the two diaphragms, and accordingly determines its thickness, strength, and the composition of the trapped gas. Preferably, for the topological event of the liquid front collision and merging, the configuration state reconstruction network further outputs whether the two oxide films are welded when they meet and the configuration of the oxide joint formed when they are not welded, so as to characterize the unwelded oxide joint formed when the thin-walled part or multi-gate part is filled.

[0015] Specifically, in step five, the entangled membrane with its configuration determined in step four is taken as a membrane entity within the bulk phase. It undergoes Lagrangian propagation with the velocity field of the bulk melt, and its configuration is updated based on subsequent contact events during this propagation. When the solidity at the location of the membrane entity reaches a set threshold and solidifies, the membrane entity, along with its configuration, is frozen at that location. Furthermore, for membrane entities that subsequently adhere and stack, their bilayerity and spatial orientation are updated based on steps three and four to reproduce the secondary folding of the membrane within the bulk phase. This deepens the hazard determination from a single static assessment at the moment of entanglement to a path-dependent result determined by the dynamics of the entire process of filling, holding pressure, and solidification.

[0016] Specifically, in step six, the configuration state of the film entities frozen at various locations in the casting is statistically analyzed, and a harmful double-layer film distribution field is generated inside the casting. This distribution field outputs the density, spatial orientation, and maturity of the harmful double-layer film at each spatial location. Then, the harmful double-layer film distribution field is combined with the stress field, temperature field, and solidity field at that spatial location to predict the tendency for thermal cracking, porosity, and fatigue crack initiation at that spatial location. The tendency for fatigue crack initiation is determined based on the density of the harmful double-layer film, the angle between its spatial orientation and the principal stress direction at that spatial location, and its maturity.

[0017] Furthermore, the method of the present invention also includes step seven, using pouring temperature, holding time, gate position, and filling speed as variables to be optimized, and the distribution of harmful double-layer film and casting defect tendency predicted in step six as optimization targets, iteratively adjusting the variables to be optimized within the process parameter space and repeating steps one to six until the distribution of harmful double-layer film and casting defect tendency meet the set requirements, and outputting the optimized process parameters. Preferably, the gate position and pouring temperature are adjusted with the aim of lower filling speed, stable liquid surface, and fewer events of self-contact of free surfaces and collision and merging of liquid fronts, so as to suppress the formation of harmful double-layer film.

[0018] In another aspect, the present invention provides an aluminum alloy melt casting simulation system, comprising: a basic physics field solving module, used to solve the basic physics field of the casting and solidification process, and outputting the geometric position of the free surface, the velocity and temperature of the melt at the free surface, and the velocity field and solidity field of the bulk melt step by step; a surface material entity graph construction module, used to discretize the free surface into surface material entities corresponding to a diaphragm, construct a dynamic surface material entity graph with each surface material entity as a node and the connectivity between adjacent diaphragms as edges, wherein the nodes carry the position of the diaphragm, the unit normal vector, the cumulative exposure time since exposure to the atmosphere, and features identifying the wetting state of the two diaphragm surfaces, and create new surface material entities at newly generated free surfaces, so that the individual identity and folding topology of the diaphragms continue across time steps; and a topology event recognition module, used to input the dynamic surface material entity graph. The system comprises a graph neural network for topological event recognition, which outputs the topological event types and participating diaphragm pairs via message passing, including free surface self-folding and bonding, liquid front collision and merging, and injection fracture and splashing. A configuration state reconstruction module is used to reconstruct the configuration state network from the diaphragm pairs participating in each topological event and their contact geometry inputs, outputting the configuration state of the formed rolled-in diaphragm (whether it is a double-layer or single-layer film, whether the two surfaces of the double-layer film are welded, its spatial orientation relative to the casting coordinate system, and its maturity. A diaphragm tracking and freezing module is used to treat the rolled-in diaphragm with the configuration state as an internal film entity within the bulk phase, and to perform Lagrangian propagation with the velocity field of the bulk melt, freezing it along with the configuration state when the solidity at its location reaches a threshold. A defect prediction module is used to statistically analyze the configuration states of the film entities frozen at various locations in the casting, generate a harmful double-layer film distribution field, and predict casting defects by combining it with the stress field. The system also includes a process parameter optimization module, which executes step seven and outputs optimized process parameters. Among the above modules, the data processed by the topology event recognition module and the configuration state reconstruction module are physical quantities with definite technical meanings, such as free surface geometry, melt velocity, temperature, oxide film exposure time and orientation, and are based on the natural law of fluid free surface folding mechanics and aluminum alloy surface oxidation dynamics.

[0019] Compared with existing technologies, this invention explicitly represents the oxide film on the free surface of the melt as a dynamic surface material entity carrying individual identity, generation time, and folding topological relationships. It uses a topological event recognition network and a configuration state reconstruction network to reconstruct the bilayer nature, welding state, spatial orientation, and maturity of the film formed at the moment of free surface folding and liquid front merging. This overcomes the shortcomings of existing technologies, which only use scalar or tracer particles to characterize the entrained oxides and irreversibly lose the film configuration state record when topological changes occur on the free surface. Therefore, it outputs the distribution of harmful bilayer films throughout the solidified casting, realizing the translation from the distribution of the number of entrained oxides to the actual distribution of harmful bilayer films. Especially in situations where free surface contact and liquid front merging occur densely, such as turbulent casting, flow fracture, and the convergence of multiple liquid fronts in thin-walled parts and multi-gate parts, this invention can provide film configuration state results that can be used for defect identification. Based on this, casting defects can be predicted more accurately, casting process parameters can be optimized, casting quality can be improved, and the development cycle can be shortened. Attached Figure Description

[0020] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] Figure 1 This is a schematic diagram of the overall process of the aluminum alloy melt casting simulation method provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the data flow for the construction and cross-time step continuation of the dynamic surface material entity map provided in the embodiments of the present invention; Figure 3 This is a schematic diagram illustrating the structure and processing flow of the topology event recognition network and configuration state reconstruction network provided in this embodiment of the invention; Figure 4 This is a structural block diagram of the aluminum alloy melt casting simulation system provided in the embodiments of the present invention; Figure 5 This is a cross-sectional schematic diagram of the physical scenario of casting and filling of thin-walled multi-gate castings provided in an embodiment of the present invention; Figure 6 This is a schematic cross-sectional view of the spatial distribution and orientation of the harmful double-layer film inside the solidified casting provided in the embodiment of the present invention; Figure 7 This is a comparison and ablation analysis diagram of the spatial coincidence rate between the densely populated area of ​​the harmful double-layer film and the actual fatigue failure location provided in the embodiments of the present invention; Figure 8 This is a single grayscale heatmap of the fatigue crack initiation tendency index with respect to orientation factor and maturity, provided in an embodiment of the present invention. Figure 9 This is a biaxial composite graph showing the convergence of closed-loop process parameter optimization iterations provided in this embodiment of the invention. Detailed Implementation

[0022] To make the objectives and technical solutions of this invention clearer, the embodiments of this invention will be further described in detail below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of this invention and should not be construed as limiting the scope of protection of this invention.

[0023] Example 1 This embodiment provides a method for simulating aluminum alloy molten casting. Under given mold geometry, gate location, and process parameters, it reproduces the casting, filling, and solidification process of aluminum alloy molten metal, and outputs the configuration state of the oxide film entrained at various points inside the solidified casting, thereby predicting casting defects and optimizing casting process parameters. The method is generally divided into a simulation inference process driven by two trained neural networks, and a training process for obtaining these two neural networks. Steps one through seven of the simulation inference stage are described in full below, followed by the training stage of the two neural networks.

[0024] In the simulated reasoning phase, the method includes steps one through seven. Figure 1 The overall flow of the method described in this embodiment from step one to step seven is shown.

[0025] Step one involves solving the fundamental physical fields of the aluminum alloy molten metal pouring and solidification process. Specifically, using the geometric model of the mold, gate location, initial melt temperature, initial mold temperature, pouring temperature, holding time, and filling speed as inputs, the continuity, momentum, and energy equations of the melt are solved over time using the finite volume method or finite element method within the computational domain formed by the mold cavity and the gating system. The continuity and momentum equations describe the molten metal's flow within the mold cavity. The momentum equation incorporates gravity, pressure gradient, viscous stress, and the resistance to flow from the solidification of the mushy region. The energy equation describes the convective heat transfer, heat conduction, and release of latent heat of solidification, coupled with phase transitions through the relationship between melt temperature and solid fraction. The solution process is performed step-by-step, with the time step size determined according to the filling speed and mesh size based on the Courant number condition. Typical values ​​ensure that the melt advances no more than one mesh element within each time step.

[0026] In tracing the free surface of the melt, this step employs either the fluid volume method or the level set method. Taking the fluid volume method as an example, a melt volume fraction is defined on each grid cell, with a value of 1 indicating that the cell is completely filled with melt, 0 indicating that the cell is a cavity, and values ​​between 0 and 1 indicating that the cell contains a free surface. By solving the transport equation of the melt volume fraction and combining it with interface reconstruction, the geometric position and normal of the free surface can be obtained step-by-step. Therefore, this step outputs the following quantities step-by-step: the geometric position and unit normal vector of the free surface, the melt velocity and temperature at the free surface, the velocity field of the bulk melt, and the solidity field solved from the energy equation and phase transition relation. The solidity field is subsequently used to determine whether the film has solidified and frozen; the stress field is obtained in the later stages of solidification using a thermo-elastic-plastic constitutive model from the temperature field and the solidity field, and is used for subsequent defect prediction. As a specific parameter setting, the pouring temperature is typically set approximately 50 to 100 degrees Celsius above the liquidus line of the aluminum alloy. The initial mold temperature varies from room temperature to several hundred degrees Celsius depending on whether it is a metal mold or a sand mold. The filling rate is calculated by back-calculating the gate cross-section and the predetermined filling time. The latent heat of solidification in the energy equation is determined based on the phase diagram and specific heat of the aluminum alloy, and the relationship between solid fraction and temperature is given by the solidification path model used. The physical fields output by the solution are defined spatially on the same computational grid and aligned temporally to the same set of time steps. This ensures that the position, velocity, temperature, and solid fraction of each film in the subsequently constructed dynamic surface material entity map are derived from a consistent spatiotemporal discretization, providing a unified benchmark for data transfer across steps and modules.

[0027] Furthermore, preferably, while solving the aforementioned fundamental physical fields, a surface age transport equation is additionally solved on the free surface, so that each point on the free surface carries the accumulated exposure time since its formation upon exposure to the atmosphere. The surface age transport equation is, in form, a scalar transport equation with convection along the free surface and a source term: its convection term transports the existing surface age with the melt velocity at the free surface, and its source term accumulates the exposure time on the free surface by increasing the surface age by one unit per unit time, while at newly formed free surfaces, the surface age is initialized to 0. More preferably, the source term of the surface age transport equation is not simply accumulated linearly over time, but rather the film thickness growth rate is determined according to the local temperature and oxygen partial pressure of the free surface based on oxidation growth kinetics. Specifically, the film thickness growth rate is determined according to a parabolic oxidation law, and its film thickness increment is related to the temperature-dependent oxidation rate constant and the square root of the exposure time, thereby making the exposure time correspond to the actual thickness and strength of the oxide film. Therefore, oxide films formed under prolonged exposure and high-temperature conditions correspond to longer exposure times and thicker, stronger films that are more difficult to re-weld by the melt. This provides a quantity with precise physical meaning for subsequent determination of the weldability and maturity of the film. It should be noted that surface age and exposure time are two names for the same physical quantity in this article. The definitions of equivalent film thickness, normalized exposure time, and maturity derived from these terms, as well as their interconversion relationships, are uniformly given in step two below.

[0028] Step two involves constructing a dynamic surface material entity map that evolves during the filling process. Specifically, at each time step, the free surface at the current moment is geometrically discretized into a set of surface material entities. Each surface material entity corresponds to a membrane on the free surface, and the scale of the membrane can be equivalent to the interface mesh unit or adaptively subdivided according to curvature. Each surface material entity serves as a node in the dynamic surface material entity map, and the feature vector carried by the node includes at least: the spatial coordinates of the membrane, the area of ​​the membrane, the unit normal vector of the membrane, the cumulative exposure time of the membrane since its generation from exposure to the atmosphere, the local temperature at the membrane, and the wetting state indicating whether the two membrane surfaces face inward towards the melt and outward towards the air gap, respectively. The wetting state is used to distinguish the two surfaces of the membrane: the membrane surface facing the melt is usually wetted by the melt, while the membrane surface facing the air gap is a non-wetting free surface. This distinction is the basis for subsequent determination of whether the bilayer membranes are non-wetting surfaces that adhere to each other.

[0029] To avoid ambiguity caused by the substitution of various quantities related to the oxide film formation history in different paragraphs, this paper provides a unified definition for this group of quantities, all of which are derived from the same basic quantity. Firstly, exposure duration. The fundamental quantity is defined as the accumulated time at a point on a free surface from the moment it was generated by exposure to an oxygen-containing atmosphere to the present moment, with the dimension of time, and is obtained from the surface age transport equation described in step one; where surface age is the quantity that acts as a field on the free surface. The term "time" refers to the exposure duration, which is the value of the field quantity on a specific membrane. The two terms refer to the same physical quantity, have equal numerical values ​​and the same dimensions, and are interchangeable. Secondly, equivalent film thickness... The oxide film thickness is obtained by mapping the exposure time to the oxidation growth kinetics, and is taken according to the parabolic oxidation law. ,in To vary with local temperature The changing oxidation rate constant, The dimension of this is length. Thirdly, normalized exposure time. Exposure duration for reference duration The result of dimensionless transformation is taken as follows: The value ranges between 0 and 1, where The characteristic time required for the oxide film of the aluminum alloy to regrow at the casting temperature until it is difficult to be re-welded by the melt is taken as the fourth characteristic. This is a dimensionless comprehensive index with values ​​between 0 and 1, characterizing the thickness, strength, and difficulty of re-welding the film by the melt. It is obtained from the oxidation growth kinetics based on the exposure time and local temperature. ,in This corresponds to the reference film thickness in the non-weldable state. From the above definitions, it can be seen that the equivalent film thickness is a monotonic mapping of the exposure time through oxidation kinetics, the normalized exposure time is the dimensionless result of the exposure time, and the maturity is the result of normalizing the equivalent film thickness to the reference film thickness. Under the isothermal approximation where the oxidation rate constant can be considered constant, the three satisfy... Therefore, maturity, equivalent membrane thickness, and normalized exposure duration are monotonically one-to-one related. In cases where only their monotonic dependence is used without involving specific dimensions, any of them can be used to represent them. In all other cases, the above definition shall be used, and different quantities shall not share the same symbol.

[0030] To ensure that a single polarity is sufficient to uniquely determine the wetting state of each of the two membrane surfaces, this paper makes the following stipulations regarding the normal direction of the membrane, the numbering of the two membrane surfaces, and the transformation rules during flipping. Firstly, the normal direction stipulation: for a membrane located on a free surface, its unit normal vector... The first rule is to use the outward normal vector pointing from the melt side to the gas phase side. The second rule is the film surface numbering rule: based on this unit normal vector, [the film surface is numbered]. The side of the membrane that is pointed to is designated as the front side. The membrane surface it points to is denoted as negative. Thirdly, there are rules for determining the values ​​of polar quantities: polar quantities... Indicate the wetting status of the two membrane surfaces. Pick or , This indicates that the front side is the non-wetting side and the back side is the wettable side. This indicates that the front side is the wettable surface and the back side is the non-wetting surface; for a newly formed film on a free surface, its outer normal side is the gas phase side, therefore its polarity is always initialized to 0. Fourthly, the reversal rule: when the recording direction of the normal of the diaphragm is changed to the opposite direction due to folding of the free surface, being drawn into the melt, or subsequent movement, the numbers of the two diaphragm surfaces are interchanged, and the polarity quantities are simultaneously reversed, i.e. Transform into Since the normal and polarity quantities transform in pairs according to the above rules, the wetting state of each of the two membrane surfaces is uniquely determined under any normal recording method, and is independent of the choice of the positive or negative direction of the normal. Therefore, a combination of a unit normal vector and a single polarity quantity is sufficient to completely express the state of the two membrane surfaces. Accordingly, when the two membranes are attached together, the membrane... The mating surface is the vector from its centroid to the opposite square's center and The dot product is determined as follows: when the dot product is positive, the mating surface is the front side; when the dot product is negative, the mating surface is the back side. Then, the polarity is determined... This allows us to determine whether the bonding surface is a non-wetting surface. Only when both bonding surfaces of the two films are non-wetting surfaces is the necessary condition for forming a bilayer film met.

[0031] Between nodes, adjacent membranes are connected by edges, which represent the connectivity of the membranes on the free surface. That is, two membranes are adjacent to each other on the free surface and share a boundary. Furthermore, the edges connecting the nodes of adjacent membranes also carry the relative orientation between the adjacent membranes, that is, the angle between the unit normal vectors of the two membranes. Thus, the dynamic surface material entity graph not only records the independent properties of each membrane, but also records the connectivity of the entire free surface in terms of graph topology.

[0032] The key to this step is ensuring the continuity of the membrane's identity across time steps. As the free surface expands due to melt propagation, new surface material entities are created at the newly generated free surface, and their creation time is recorded at the current moment, while their connectivity with existing membranes is updated. For existing membranes, their identity remains unchanged, and only their time-varying characteristics such as position, area, and exposure duration are updated. Preferably, each surface material entity is not re-discrete at each time step, but rather undergoes a Lagrange propagation with the melt velocity at the free surface. That is, the position at the current time step is obtained by multiplying the melt velocity at the free surface by the time step length, starting from the position of the membrane at the previous time step. The area of ​​the membrane is updated according to the local stretching rate of the free surface. When the free surface is stretched at a certain point, the area of ​​the membrane on it increases accordingly while the membrane thickness decreases accordingly, and vice versa when it is compressed. Thus, a Lagrangian description of free surface materials is superimposed on the Eulerian description of bulk melt, allowing the individual identity, generation time, and folding topological relationship of each membrane to continue across time steps. This is the prerequisite for reconstructing the folding configuration of the membrane. Figure 2 The data flow of the above dynamic surface material entity map being constructed, advanced, and continued between two adjacent time steps is shown.

[0033] Step 3: Identify topological events on the free surface. Specifically, the dynamic surface material entity graph obtained in Step 2 is input into a topological event identification network, which is a graph neural network. The graph neural network takes the node and edge features of the graph as input. Through multiple rounds of message passing on the surface material entity graph, in each round, each node aggregates the transformed features of its neighboring nodes and updates its own hidden state. This ensures that the representation of each membrane includes not only its own geometric and kinematic features but also the connectivity and folding information of its neighborhood. After several rounds of message passing, the network outputs the topological events occurring or about to occur on the free surface and their types at the node and node pair levels. These types include at least three categories: free surface self-folding and merging, i.e., the same connected free surface folds and its two parts come close together; two propulsion fluid fronts collide and merge, i.e., two originally unconnected free surfaces propel each other towards each other and merge; and flow fracture and splashing, i.e., the flow fractures and splashes during its descent, creating new isolated free surfaces. For each identified event, the network also outputs the diaphragm pair involved in the event and its contact geometry, the specific parameters of which are described in step four below.

[0034] The geometric and kinematic features upon which the network is based include at least the following: the distance between two closely approaching membranes, the angle between their unit normal vectors, their relative approach velocity, and their respective exposure durations. These features characterize whether the two membranes are about to contact, whether the contact is face-to-face or in another manner, the intensity of the contact, and the maturity of their respective oxide films. Preferably, to improve the detection rate and reduce the network load, this step performs a candidate screening based on geometric criteria before inputting the graph into the network: for pairs of closely approaching membranes, when their unit normal vectors are opposite (i.e., the two membrane surfaces are facing each other) and the distance is less than a set threshold, they are determined to be candidate events of self-folding and bonding or liquid front collision and merging. Only the subgraphs corresponding to these candidate events are submitted to the topological event recognition network for further classification. Since there are numerous candidate events in situations where self-contact events occur intensively, such as turbulent casting and multiple liquid front convergence, this two-stage processing of geometric screening followed by network classification can control the computational load while ensuring the detection rate.

[0035] Step four: Reconstruct the configuration state of the membranes being rolled up. Specifically, for each topological event identified in step three, a reconstruction network is built for the membrane pairs involved in the event and their contact geometry input configuration states. The network outputs the configuration state of the membranes rolled up by the event.

[0036] The contact geometry is a set of quantities describing the contact state of two diaphragms participating in the same topological event at the moment the event occurs. Specifically, it includes the following parameters: the position coordinates of the contact point in the casting coordinate system; the unit normal vector of the contact surface, i.e., the normal of the interface between the two diaphragms; the distance between the centroids of the two diaphragms; the angle between the unit normal vectors of the two diaphragms, which, according to the orientation convention described later in this step, is the absolute value of the cosine or the angle folded to the range of 0 to 90 degrees; the relative approach velocity of the two diaphragms, i.e., the projection of the difference in velocity between the two diaphragms along the direction of the line connecting their centroids; the overlapping contact area of ​​the two diaphragms, i.e., the area of ​​the overlapping portion of the two diaphragms after projection along the normal of the contact surface; and the number and polarity of the respective contact surfaces of the two diaphragms, used to determine whether both contact surfaces are non-wetting surfaces according to the rules agreed upon in step two. Among the above parameters, the spacing, the included angle of normal, and the relative approach speed are the geometric and kinematic features used for event recognition in step three. The contact point position and the contact surface normal are given by the regression of the topological event recognition network in step three on the positive sample membrane pair. The overlapping contact area is calculated by the area of ​​the two membranes and their projection overlap relationship.

[0037] The configuration state includes at least four aspects: whether the wound diaphragm is a double-layer diaphragm or a single-layer diaphragm, whether the two surfaces of the double-layer diaphragm are welded together, the spatial orientation of the diaphragm relative to the casting coordinate system, and the maturity of the diaphragm.

[0038] When determining the bilayer and welded state, the configuration state reconstruction network makes a comprehensive judgment based on whether the membrane surfaces of the two membranes involved in the event are both non-wetting membrane surfaces, whether there is an air gap in the middle during the bonding, and the strength and weldability of the membrane determined by the exposure time of each membrane: when the two membranes are bonded together with their non-wetting membrane surfaces, with an air gap in the middle, and the exposure time of the two membranes is relatively long and the membrane is relatively thick and strong, making it difficult to be re-welded by the melt, the formed membrane is determined to be a harmful bilayer unwelded air-folded state; conversely, when the bonding is a membrane surface that can be wetted by the melt, or the membrane is relatively thin and new and can be re-welded by the surrounding melt, it is determined to be a harmless single-layer or welded configuration. When determining the spatial orientation, this paper consistently uses the unit normal vector of the diaphragm surface in the casting coordinate system to represent its spatial orientation, instead of the plane in which the diaphragm lies. The plane in which the diaphragm lies is the plane with the unit normal vector as its normal direction; the two are mutually unique and equivalent expressions. Wherever the plane in which the diaphragm lies is mentioned, its spatial orientation is calculated according to the unit normal vector of that plane. Furthermore, since the diaphragm surface is a curved surface with two sides, its normal direction exists... and The two directions represent the same spatial orientation, meaning the spatial orientation is based on an undirected axis rather than a directed vector. Therefore, all calculations involving the included angle, similarity, or loss of the membrane's spatial orientation should use the absolute value of the dot product of the two unit normal vectors. Or, equivalently, fold the included angle between the two to the range of 0 to 90 degrees, based on this... and The same result is given. It should be noted that the above absolute value convention only applies to the comparison and regression of the spatial orientation attribute; when using the opposite direction of the unit normal vectors of the two diaphragms as the criterion for candidate events in step three, it is based on the external normal vector from the melt to the gas phase agreed upon in step two. The direction of this normal vector is uniquely determined by the relative position of the melt and the gas phase, and its positive or negative sign has a definite physical meaning, used to determine whether the two diaphragms are facing each other with their non-wetting surfaces facing each other. Therefore, the signed dot product of the normal vectors is still taken here instead of its absolute value. Based on this, the network transforms the geometric orientation of the two diaphragm contact surfaces at the moment the event occurs to the casting coordinate system, and outputs the unit normal vector of the formed diaphragm surface in the casting coordinate system. This orientation is used according to the above absolute value convention when performing angle calculations with the principal stress direction in the subsequent process. When determining maturity, the network determines the maturity of the formed membrane based on the greater of the exposure times of the two membranes, and accordingly determines its thickness, strength, and the composition of the trapped gas, because the hazard of a bilayer membrane is mainly dominated by the older, thicker, and more difficult-to-weld membrane. Preferably, for events involving the collision and merging of two propulsion fluid fronts, the configuration state reconstruction network further outputs whether the two oxide films are welded at the time of encounter and the configuration of the oxide joint formed when they are not welded, in order to characterize the unwelded oxide joints formed during the filling of thin-walled parts and multi-gate parts. The oxide seam configuration is a set of quantities describing the seam formed at the meeting point of two liquid fronts. Specifically, it includes the following parameters: bilayerity, i.e., whether the seam is formed by two oxide films adhering to each other to form a bilayer structure; the welding state of the two films, i.e., whether the two oxide films that meet have been re-welded by the melt to eliminate the interface between them, represented by a binary identifier or a probability of non-welding; the thickness of the air gap, i.e., the equivalent thickness of the gas layer sandwiched between the two films, taken as 0 when no gas is sandwiched; the orientation of the seam, i.e., the unit normal vector of the seam surface in the casting coordinate system, which is equivalent to the reverse of the normal vector according to the aforementioned orientation convention; the area of ​​the seam and its extension range in the casting coordinate system; and the larger of the maturity of the two films on both sides of the seam. The parameters mentioned above are consistent with the quantities used for the membrane entity in steps five and six. That is, the bilayer nature and welding state of the seam correspond to the bilayer nature and welding state of the membrane entity, the orientation of the seam corresponds to the spatial orientation of the membrane entity, the maturity of the seam corresponds to the maturity of the membrane entity, and the area of ​​the seam is included in the statistics of the harmful bilayer membrane area density described in step six. Figure 3 The structure and processing flow of the topology event recognition network and the configuration state reconstruction network are shown.

[0039] Step 5: Track the entangled membrane and freeze its configuration during solidification. Specifically, the entangled membrane with its configuration determined in Step 4 is considered as the membrane entity within the bulk phase. Unlike the surface material entity located on the free surface in Step 2, the membrane entity has been entangled within the melt and thus undergoes Lagrangian propagation with the velocity field of the bulk melt. That is, at each time step, its position is updated by multiplying the bulk melt velocity at the location of the membrane entity by the time step. During propagation, the membrane entity updates its configuration based on subsequent contact events: preferably, for membrane entities that subsequently adhere or stack, their bilayerity and spatial orientation are updated according to Steps 3 and 4 to reproduce the secondary folding of the membrane within the bulk phase. Similarly, the expansion and contraction of the air gap between the membrane entities under the pressure field during the holding phase, and the possibility of the membrane being re-welded by the surrounding melt, can also be continuously updated as latent states along the Lagrangian trajectory of the membrane entity. Therefore, the determination of whether a membrane will eventually develop into a harmful bilayer membrane is no longer a static judgment at the moment of entrapment, but a path-dependent result jointly determined by the dynamics of the entire process of filling, holding pressure, and solidification. When the solid fraction at the location of the membrane entity reaches a set threshold and solidification occurs, the melt has lost its fluidity at that point, and the membrane entity no longer moves with the flow. Thus, the membrane entity, along with its current configuration, is frozen at that location. The set threshold is usually taken as the solid fraction value corresponding to the dendritic coherence point or the point at which the melt loses its macroscopic fluidity. For example, for common aluminum alloys, it can be taken as a solid fraction between 50% and 70%. When the solid fraction of the grid containing the membrane entity first exceeds this threshold, freezing is triggered. Afterward, the position and configuration of the membrane entity no longer update with the flow, and only the air gap it contains can still be corrected by limited expansion and contraction according to the holding pressure.

[0040] Step Six: Output the distribution field of harmful double-layer films inside the casting and predict casting defects. Specifically, after filling and solidification, the configuration state of the film entities frozen at various locations in the casting is statistically analyzed, generating a distribution field of harmful double-layer films inside the casting. This distribution field outputs the density, spatial orientation, and maturity of the harmful double-layer films at each spatial location. The density is defined as the ratio of the sum of the film surface areas of all harmful double-layer film entities within the statistical volume element at that spatial location to the volume of that volume element, i.e., the area of ​​the harmful double-layer film per unit volume, denoted as... Its dimension is the negative first power of length. Area density is used here instead of number density because: the number of films depends on the discrete scale taken for the free surface in step two; the number changes when the same oxide film is subdivided into more films, while the film area does not change with the discrete scale; and the following weighted statistics based on orientation and maturity use the film area of ​​each film as the weight, which is consistent with area density. As an alternative, number density can also be used, denoted as... , is defined as the number of harmful bilayer membrane entities per unit volume, and its dimension is the negative cube of length; and Because they have different dimensions, they cannot be substituted for each other, nor can they share the same symbol. The dimensions in the following formulas should be used consistently throughout the entire process, and adjusted accordingly. Unless otherwise specified, all densities mentioned below refer to areal density. The spatial orientation and maturity are respectively derived from the statistical results of each harmful bilayer membrane entity within the volume element, weighted by its membrane surface area, i.e., the area-weighted distribution of orientation and its principal direction, and the area-weighted average of maturity. Furthermore, the distribution field of the harmful bilayer membrane is combined with the stress field, temperature field, and solidity field of that spatial location to predict the tendency for thermal cracking, porosity, and fatigue crack initiation at that location. Specifically, the tendency for thermal cracking increases mainly in the later stages of solidification where there is high solidity, high tensile stress, and a relatively high density of the harmful bilayer membrane; the tendency for porosity increases mainly where the harmful bilayer membrane encloses air gaps and where solidification feeding is insufficient; and the tendency for fatigue crack initiation is determined by the density of the harmful bilayer membrane, the angle between its spatial orientation and the principal stress direction at that spatial location, and its maturity. Specifically, when the bilayer membrane density is high, its membrane surface normal is nearly parallel to the principal stress direction (i.e., the membrane surface is perpendicular to the principal stress), and the membrane is relatively mature and thick, the tendency for crack initiation under fatigue load is highest at that location. As a specific implementation, the aforementioned tendency for fatigue crack initiation can be quantified as a scalar index, which is obtained by multiplying three factors: the first is the density factor, which is the area density of the harmful bilayer film at that location. The ratio of the density to the reference area makes this factor dimensionless; the higher the density, the higher the exponent. The second factor is the orientation factor, which is the square of the cosine of the angle between the film surface normal and the principal stress direction at that location. The orientation factor reaches its maximum value when the membrane surface is perpendicular to the principal stress, i.e., the normal direction is parallel to the principal stress. When the membrane surface is parallel to the principal stress, the orientation factor tends to 0. Since the square of the cosine is taken here, the membrane surface normal direction... and The same orientation factor is given, consistent with the positive and negative equivalence convention of orientation described in step four; the third is the maturity factor, which is taken as the maturity of the harmful bilayer membrane at that location. The monotonically increasing function, according to the conversion relationship given in step two, can also be equivalently expressed as equivalent film thickness or normalized exposure time, and the three are monotonically one-to-one corresponding to each other. The dependence of the fatigue crack initiation tendency index on the orientation factor and maturity factor is as follows: Figure 8 As shown, Figure 8 This is a single grayscale heatmap, with the horizontal axis representing the orientation factor. The vertical axis represents the normalized maturity level, and the grayscale scale represents the normalized fatigue crack initiation tendency index. The values ​​marked at both ends of the horizontal axis, indicating whether the film surface is parallel to the principal stress or perpendicular to the principal stress, represent the physical meaning of the values ​​at both ends of the same horizontal axis. The dashed lines in the figure are contour lines of the tendency index, and the numbers 0.2, 0.4, 0.6, and 0.8 noted next to them are the index values ​​of the corresponding contour lines, not the coordinate axis scale. As can be seen from the figure, the tendency index is at its maximum when both the orientation factor and maturity level increase simultaneously; this area corresponds to the fatigue weak zone. The fatigue tendency scalar index obtained in this way is distributed point-by-point inside the casting, forming a fatigue crack initiation tendency map, which can be directly used to delineate fatigue weak zones and quantitatively compare with the measured fracture location and its orientation.

[0041] Step 7: Optimize the casting process parameters. Specifically, using process parameters such as casting temperature, holding time, gate position, and filling speed as variables to be optimized, and the distribution of harmful double-layer film and the tendency of casting defects predicted in Step 6 as optimization targets, iteratively adjust the process parameters within the process parameter space and repeat Steps 1 to 6 until the distribution of harmful double-layer film and the tendency of casting defects meet the set requirements, and output the optimized process parameters. The iteration can be carried out within the process parameter space using methods such as grid search, genetic algorithm, or Bayesian optimization. Preferably, the gate position and casting temperature are adjusted with the aim of lower filling speed, stable liquid surface, and fewer events of self-contact of free surfaces and collision and merging of liquid fronts. For example, the gate height can be appropriately reduced to reduce the flow drop, and overflow and slag collection structures can be added to guide the early film-containing melt to flow out of the cavity, thereby inhibiting the formation of harmful double-layer film from the source.

[0042] The above describes the complete process of the simulation inference phase. The following describes the training phase of the topology event recognition network and the configuration state reconstruction network. Both networks are trained in a supervised manner, with training samples generated by high-fidelity forward simulation. Specifically, in the training data preparation phase, forward simulations are performed on a batch of casting examples with different mold geometry, gating arrangements, and process parameters using fundamental physics field solving and high-resolution interface tracking. The mesh resolution of the high-resolution interface tracking is sufficient to resolve the details of the folding and merging of free surfaces. For each example, the geometry of the free surface, the connectivity of the membranes, and their kinematic quantities are recorded step-by-step. Based on the actual folding and merging geometry of the interface at high resolution, the type of topology event occurring at each location, the membrane pairs involved in the event, and whether the resulting membranes are double-layered, welded, and their orientation and maturity are labeled, thus forming a labeled sample set. In the data preprocessing phase, the position, velocity, area, and exposure duration of node features are normalized according to their dimensions, and node pairs that are too far apart in the graph structure and cannot interact are pruned to control the size of the graph.

[0043] During the model training phase, the topology event recognition network and the configuration state reconstruction network can be trained separately or jointly. Taking separate training as an example, the topology event recognition network adopts a graph neural network structure, with 3 to 6 rounds of message passing and 64 to 256 hidden feature dimensions for each node. Its output is the multi-class probability of the event type and the membrane pairing; its loss function is the weighted sum of the cross-entropy loss of the event type and the matching loss of the membrane pairing. The configuration state reconstruction network takes the membrane pair features and contact geometry involved in the event as input, and outputs the classification probability of bilayerity and welding state, the regression value of orientation, and the regression value of maturity; its loss function is the weighted sum of the cross-entropy loss of the classification term and the mean squared error loss of the regression term. Training uses an adaptive moment estimation optimizer, and the initial learning rate is preferably set to... And a cosine annealing strategy is used to gradually decay to approximately during training. The optimal batch size is 16 to 64 event subgraphs per batch; the optimal number of training rounds is 120 to 300 rounds, or the convergence criterion is that the event detection rate and configuration determination accuracy on the validation set no longer improve for several consecutive rounds. During training, to alleviate the problem of imbalanced sample sizes for different event types, a larger weight can be assigned to the loss of rare event types. The two networks obtained through the above training can then be used in the aforementioned simulation inference stage to identify topological events and reconstruct the membrane configuration state step-by-step for new casting examples.

[0044] Combining the above-mentioned simulation inference and training phases, the overall data flow of the method described in this embodiment is as follows: During the training phase, the parameters of the two networks are completed offline and solidified. During the inference phase, steps one through five are executed sequentially for each new casting example, time-step by time. After filling and solidification, step six is ​​executed to output the distribution field of harmful double-layer films and defect tendencies, and optionally, the process parameter optimization loop in step seven is entered. The two neural networks are only invoked in steps three and four; the remaining steps are conventional physics field solving, graph structure maintenance, and Lagrange tracing. Therefore, the method can be implemented by plugging into the surface material entity graph construction, topological event recognition, and configuration state reconstruction stages on the existing casting simulation software's solution framework. This facilitates engineering implementation and allows for iterative updates of the two networks as training samples accumulate.

[0045] Example 2 In a preferred embodiment of the present invention, the construction of the dynamic surface material entity map and the topological event recognition and configuration state reconstruction driven by it are further refined in order to stably restore the folding configuration of the film under the condition that the contact of the free surface itself and the merging of liquid fronts occur densely.

[0046] In constructing the dynamic surface material entity graph, the free surface at the current time step is geometrically discretized into a set of membranes, with each membrane serving as a node in the graph. The node's eigenvector is implemented as a multidimensional vector, its components including at least the three-dimensional coordinates of the membrane centroid, the membrane area, three components of the membrane's unit normal vector, the cumulative exposure time since the membrane was exposed to the atmosphere, the local temperature at the membrane, and a polarity quantity used to identify the wetting state of the two membrane surfaces. This polarity quantity takes two signs, positive and negative. Its meaning, membrane numbering rules, and the inversion rules for its pairwise transformation with the unit normal vector are all executed according to the stipulations in step two. That is, this polarity quantity and the membrane's unit normal vector uniquely determine the wetting state of each of its two membrane surfaces. The edges of the graph are established when two membranes are adjacent to each other on the free surface. The edge features include at least the cosine of the angle between the centroids of the two membranes and the angle between the unit normal vectors of the two membranes. To ensure the membrane identity persists across time steps, each membrane is assigned an identity identifier created at the time of its generation and deactivated only after solidification or re-welding by the melt, and its generation time is recorded. Between adjacent time steps, the attributes of the membrane from the previous time step are transferred to the same membrane in the current time step using the identity identifier as an index. Only its position is updated according to Lagrange progression, its area is updated according to local stretching rate, and its exposure time is updated according to the cumulative exposure time rule, while its identity identifier and generation time remain unchanged.

[0047] In the topological event recognition network, a graph neural network with edge features is employed. The network first embeds the original node features and edge features into the latent space using a multilayer perceptron. Then, multiple rounds of message passing are performed. In each round, for each edge, the latent features of its two endpoints are concatenated with the latent features of that edge, and a message function is used to calculate a message. Each node aggregates messages from all its neighboring edges according to attention weights and updates them with its current latent features using an update function to obtain new latent features. The attention weights are calculated based on the relative approach speed and normal angle between the two membranes, giving greater weight to messages between membrane pairs approaching in opposite directions. The number of message passing rounds is preferably 3 to 6, the node latent feature dimension is preferably 64 to 256, and the number of attention heads is preferably 4 to 8. After message passing, the network outputs the probability of each candidate membrane pair experiencing self-folding and adhesion, liquid front collision and merging, or injection breakage and splashing at the node pair level, and regresses the position of the contact point and the normal of the contact surface on positive sample pairs.

[0048] In terms of the configuration state reconstruction network, the latent features of the two membranes involved in the event after being encoded by the above-mentioned graph neural network, the exposure time of the two membranes, the polarity of the contact surface of the two membranes, and the contact geometry are used as inputs. The multilayer perceptron outputs four sets of quantities: the first is the binary probability of whether the involved membrane is a double membrane or a single membrane; the second is the binary probability of whether the two membrane surfaces are welded when it is determined to be a double membrane; the third is the spatial orientation of the formed membrane in the casting coordinate system, which is represented by the unit normal vector of its membrane surface and regressed. According to the convention described in step four, the unit normal vector and its opposite direction represent the same orientation; the fourth is the maturity of the membrane. Its value is between 0 and 1, and regression is performed. Based on the conversion relationship given in step two, its equivalent membrane thickness or normalized exposure time can also be equivalently regressed. To simultaneously constrain the above multiple outputs during training, a composite loss function is used, the expression of which is: ,in and The cross-entropy loss is due to the dual-layer nature and the welded state. The vector regression loss for orientation is taken as 1 minus the absolute value of the dot product of the predicted normal and the true normal, i.e. The absolute value is used here because the positive and negative directions of the membrane surface normal represent the same spatial orientation. If a regular dot product were used, it would... and Treating them as different orientations leads to erroneous penalties imposed on the network due to the arbitrariness of the way positive and negative normals are recorded. The mean squared error loss for maturity. to The weighting coefficients for each item are selected, with the preferred values ​​being... and The weight of the regression term is greater than that of the regression terms of orientation and maturity, so as to prioritize the accuracy of the key judgment of whether it is harmful or not.

[0049] In terms of beneficial effects, by employing the aforementioned dynamic surface material entity map with identity identifiers, combined with a graph neural network featuring edge features and attention, and a multi-output configuration state reconstruction network, the connectivity and folding topology of the membrane can be preserved when the free surface undergoes folding and merging. Therefore, the output is no longer a grid-by-grid scalar of entrainment or isolated particle position, but rather a membrane configuration state with bilayerity, welded state, orientation, and maturity. In turbulent casting and multi-liquid-front confluence scenarios with dense self-contact events, the event detection is relatively stable due to the two-stage processing of event recognition—first geometrically screening candidates and then classifying them by the network. The configuration determination is also more sensitive to small relative orientation differences between membranes, thus ensuring that the output of harmful bilayer membrane distributions still has the discriminative power for defect identification even in these difficult conditions. Compared to methods that characterize entrained oxides with scalars or passive particles, this embodiment provides defect prediction basis with orientation and maturity discrimination with only a limited increase in computational load.

[0050] In terms of principle analysis, the reason why the above-mentioned effects can be achieved is that this embodiment changes the way free surfaces are described. Existing methods treat free surfaces as the geometric interface of a bulk melt at a certain moment. The interface is recalculated at the next moment, and the correspondence between any point on it and the previous moment is not preserved. Therefore, the connectivity relationships changed by folding and merging cannot be recorded. This embodiment, however, establishes a cross-time step correspondence for the membranes using identity identifiers, making the free surfaces a family of material entities with individual histories. Their connectivity relationships are explicitly preserved as edges in a graph. When two membranes are bonded together, it is precisely these preserved quantities—normal, polarity, and exposure time—that determine whether the bonding forms a harmful configuration of non-wetting surface to non-wetting surface, gas trapped in the middle, and difficult to weld. The message passing mechanism of the graph neural network corresponds precisely to the physical picture of the membranes influencing each other through connectivity relationships, enabling the network to judge local folding events based on neighborhood information. It should be noted that when the resolution of the interface forward modeling simulation is insufficient to resolve extremely fine membranes, or when the proportion of a certain type of event in the training samples is too low, the accuracy of the above method in identifying and determining the configuration of such events may be affected. In this case, it is necessary to improve the forward modeling resolution or supplement the corresponding samples.

[0051] Example 3 In another embodiment of the present invention, as an optional alternative, instead of using graph neural networks to identify events in the membrane diagram at each time step, the reconstruction of the membrane configuration state is transformed into a time-series modeling problem along the Lagrange trajectory of the membrane entity, so as to further characterize the evolution of the configuration state of the membrane from its self-generation to its involvement in the solidification and freezing process.

[0052] In this alternative approach, the fundamental physical fields are still solved and surface material entities carrying identification are constructed in steps one and two. However, instead of explicitly identifying topological events at each time step, a state trajectory is maintained for each membrane entity over time. For each membrane entity, its position, velocity, pressure and temperature of the surrounding melt, local solidity, minimum distance to its nearest neighbor membrane entities, and relative normal are used as time-step observation sequences and input into a temporal reconstruction network composed of a recurrent neural network or a temporal attention network. This temporal reconstruction network maintains a hidden state that is updated with the trajectory. The hidden state encodes information such as whether the membrane entity is currently involved, whether it is in double-layer bonding, the size of the air gap, and whether it has been re-welded by the surrounding melt. At each time step, the network updates the hidden state based on the current observation and the hidden state of the previous time step, and outputs the current configuration state of the membrane entity. When the distance between the two membrane entities is continuously less than the threshold and their relative normals are continuously opposite within several consecutive time steps, the temporal reconstruction network records a bonding in the hidden state accordingly, and judges whether it is a harmful double-layer unbonded state or a bondable state based on the exposure time of the two and the air gap between them. When the air gap between them is compressed to disappear during the subsequent pressure holding stage when the pressure increases, or when the membrane entities remain at high temperature for a sufficient period of time, the network can flip its hidden state from unbonded to bonded, thereby reproducing the dynamic changes of harmfulness along the trajectory.

[0053] The training of this alternative scheme also uses forward simulation of high-resolution interface tracking as the truth source, but the labels are given in the form of a time-step configuration state sequence. The loss function is calculated at each time step for the classification and regression output of the configuration state and summed along the trajectory. The difference between this scheme and Embodiments 1 and 2 is that its core is no longer to classify events on the membrane map at a certain time step, but to perform time-series inference of the configuration state of each membrane entity across the entire trajectory. Therefore, it has better adaptability in thick-walled parts and long-holding-pressure processes where air gap expansion and contraction and secondary welding are more significant during the pressure holding stage. The cost is that it is necessary to maintain and advance the hidden states of a large number of membrane entities, and the computational and storage overhead increases with the number of membrane entities. As an optional implementation, the time-series reconstruction network of this embodiment can also be combined with the topology event recognition network of Embodiment 1. When the latter detects a bonding event, it triggers the former to start time-step configuration state updates for the relevant membrane entities, thereby balancing the efficiency of event detection and the complete characterization of configuration evolution.

[0054] Example 4 In one embodiment of the present invention, an aluminum alloy melt casting simulation system is provided for implementing the methods described in the foregoing embodiments. The system includes a basic physics field solving module, a surface material entity map construction module, a topological event recognition module, a configuration state reconstruction module, a diaphragm tracking and freezing module, a defect prediction module, and a process parameter optimization module. The functions of each module correspond one-to-one with the aforementioned method steps. Figure 4The structural block diagram of the system is shown.

[0055] The fundamental physics field solution module is used to solve the fundamental physics fields of the casting, filling, and solidification processes. It outputs the geometric position of the free surface, the velocity and temperature of the melt at the free surface, and the velocity and solidity fields of the bulk melt step by step. This module takes the mold geometry, gate position, and process parameters as input. Its output of free surface geometry and kinematic quantities is sent to the surface material solid model construction module, its output of bulk velocity and solidity fields is sent to the diaphragm tracking and freezing module, and its output of solidity and temperature fields is used to derive the stress field before being sent to the defect prediction module.

[0056] The surface material entity graph construction module discretizes the free surface into surface material entities, each corresponding to a membrane. A dynamic surface material entity graph is constructed using these entities as nodes and the connectivity between adjacent membranes as edges. Each node carries the membrane's position, unit normal vector, cumulative exposure time since exposure to the atmosphere, and features identifying the wetting states of the two membrane surfaces. New surface material entities are created at newly generated free surfaces, ensuring the individual identity of the membranes and their folding topological relationships persist across time steps. This module maintains and outputs the dynamic surface material entity graph step-by-step, which is then sent to the topology event recognition module.

[0057] The topology event recognition module is used to input the dynamic surface material entity map into a graph neural network for topology event recognition. After message passing, it outputs the topology event types, such as free surface self-folding and bonding, liquid front collision and merging, and injection fracture and splashing, as well as the membrane pairs involved in the events. This module sends the identified events and membrane pairs, along with their contact geometry, to the configuration state reconstruction module.

[0058] The configuration state reconstruction module is used to reconstruct the configuration state network of the membrane pairs participating in each topological event and their contact geometry inputs, and outputs the configuration state of the formed wound membrane as either a double-layer membrane or a single-layer membrane, whether the two membrane surfaces of the double-layer membrane are welded together, its spatial orientation relative to the casting coordinate system, and its maturity. This module sends the wound membrane with the configuration state to the membrane tracking and freezing module.

[0059] The diaphragm tracking and freezing module is used to treat the wound diaphragm with the aforementioned configuration as an internal film entity within the bulk phase. This film entity is Lagrangian-propelled along with the velocity field of the bulk melt, and frozen along with its configuration state when the solidity at its location reaches a threshold. This module then sends the film entities frozen at various locations on the casting, along with their configuration states, to the defect prediction module.

[0060] The defect prediction module is used to statistically analyze the configuration state of the film entities frozen at various locations in the casting, generate a harmful double-layer film distribution field, and predict casting defects by combining the stress field. The process parameter optimization module is used to iteratively adjust the process parameters with pouring temperature, holding time, gate position, and filling speed as variables to be optimized, and the distribution of harmful double-layer films and the tendency of casting defects as optimization objectives, and triggers the above modules to run again, outputting the optimized process parameters.

[0061] The modules described above can be deployed on the same computing server, or the basic physics field solving module can be deployed on a high-performance computing node with parallel computing capabilities, while the topology event recognition module and configuration state reconstruction module can be deployed on a computing node equipped with a graphics processing unit. The modules exchange time-step physics field and graph data through data interfaces. Among these modules, the data processed by the topology event recognition module and the configuration state reconstruction module are all physical quantities with definite technical meanings, corresponding to free surface geometry, melt velocity, temperature, oxide film exposure time and orientation, etc.

[0062] Example 5 To further illustrate the implementation process and technical effects of this invention, an application example is given below. This example involves the casting of a thin-walled, multi-gating aluminum alloy structural component. The component has a thin wall thickness and multiple gating gates. During filling, multiple liquid fronts converge within the mold cavity, presenting a challenging situation where contact between the free surface and the collision of liquid fronts occur simultaneously. The physical scenario of this casting and filling process is as follows: Figure 5 As shown, the molten metal from each gate is injected from the bottom of the cavity and propagates upwards along the cavity. Multiple liquid fronts converge in the cavity area between adjacent gates, and the free surface exhibits turbulent rolling, drawing the oxide film into the interior of the molten metal at the convergence and folding points. In such cases, conventional simulation methods that characterize the entrained oxide using entrainment scalars or passive tracer particles are insufficient to determine whether unbonded, harmful double-layer film seams have formed at the convergence points of the liquid fronts. Therefore, their predictive ability for fatigue-weak areas in this casting is limited.

[0063] Using the method described in this invention, the basic physics field solving module first solves the filling and solidification process of the part under a given pouring temperature and filling rate, constructing a dynamic surface material entity map step by step. The topology event recognition module and configuration state reconstruction module then reconstruct the bilayer nature, welding state, and orientation of the films formed at each liquid front intersection. Next, the film tracking and freezing module freezes these films inside the casting. Finally, the defect prediction module outputs a harmful bilayer film distribution field and a fatigue crack initiation tendency map. The prediction results are compared with the defect distribution and failure locations obtained from nondestructive testing and fatigue testing after actual pouring of the part to evaluate the performance of each method.

[0064] In practical implementation, the minimum wall thickness of this thin-walled multi-gating component is approximately 2 to 3 millimeters, with three side gates. The pouring temperature is approximately 700 degrees Celsius, and the filling time is on the order of seconds. The finite volume method is used to solve the basic physics field, and the computational mesh is refined near the cavity wall to analyze the filling flow within the thin wall. The time step is dynamically determined according to the Courant number condition. The scale of the diaphragm in the dynamic surface material entity diagram is taken to be equivalent to the interface mesh unit. The identity of the diaphragm is created, advanced, and deregistered step by step. The number of diaphragm entities tracked in a single filling simulation is on the order of hundreds of thousands. After the topology event recognition network and configuration state reconstruction network are trained offline on a sample set containing dozens of examples with different gate arrangements and process parameters, they are embedded into the above time-step solution process in an inference manner. The single-step inference time is much smaller than the single-step time of the basic physics field solution. Therefore, the overall computational cost after introducing the two networks is only slightly increased compared to conventional filling simulation.

[0065] Regarding the spatial agreement between the densely populated area of ​​the harmful bilayer film and the actual fatigue failure location, the comparison results between the method of this invention and various comparative and ablation schemes are as follows: Figure 7 As shown, the overlap rate between the predicted high-probability zone and the actual failure location using the method of this invention is approximately 87% to 91%, while the overlap rate using the incorporation scalar method is approximately 62% to 68%, and the overlap rate using the passive tracer particle method is approximately 65% ​​to 70%. Regarding the orientation discrimination of fatigue weak zones, the method of this invention can provide the spatial distribution of the harmful double-layer film inside the solidified casting and the orientation of the film surface normal to the principal stress direction, such as... Figure 6 As shown, high-risk oriented films densely distributed in the liquid front convergence area between adjacent gates and with the film surface normal approaching the principal stress direction constitute fatigue weak areas. The angle distribution of the film surface normal relative to the principal stress direction given by the method of this invention has a median deviation of approximately 14 to 19 degrees from the orientation of the dominant crack surface obtained from actual fracture analysis, while neither of the two comparative schemes can output orientation information. To examine the contribution of each innovative step, an ablation comparison was conducted: when the configuration state reconstruction network was removed, only event recognition was retained, and all involved films were considered harmful, the spatial matching rate decreased to approximately 73% to 77%, and orientation information was lost; when the identity continuation mechanism of the dynamic surface material entity map was further removed and degenerated into independent discrete time steps, the event detection rate decreased significantly, and the missed detection in the dense convergence area further reduced the spatial matching rate to approximately 66% to 70%. The above data fluctuate to some extent under different mesh resolutions and different process parameters, but the advantage trend of the method of this invention over the comparative scheme is stable. Based on the prediction results, after appropriately reducing the filling speed and adjusting the gate position through the process parameter optimization module and resimulating, the density of the harmful double-layer film in the high-tendency zone of the part was reduced, indicating that the method can provide a basis for process parameter optimization.

[0066] Furthermore, based on the above prediction results, closed-loop process parameter optimization was implemented for this part: the peak value of the harmful double-layer film density in the high fatigue tendency zone output by the defect prediction module was used as the optimization target, and the pouring temperature, filling speed, and opening sequence of the three gates were used as the variables to be optimized. Bayesian optimization was employed iteratively within the process parameter space. The optimization process tended to select parameter combinations that made the convergence of multiple liquid fronts in the cavity smoother and the liquid level rise more stable, such as appropriately reducing the filling speed and staggering the opening times of each gate to avoid the simultaneous high-speed convergence of multiple liquid fronts. The re-simulation results showed that after optimization, the peak value of the harmful double-layer film density in the original high fatigue tendency zone decreased by about 30%, and the spatial range of the high fatigue tendency zone was significantly reduced, verifying the applicability of the method in closed-loop process optimization. The convergence process of the closed-loop process parameter optimization with the number of iterations is as follows: Figure 9 As shown, with the increase of the number of Bayesian optimization iterations, the peak density of harmful bilayer membranes in the high-prone region monotonically decreases from the initial value and converges to about 70%, and the spatial range of the high-prone region also decreases from the initial value and converges to about 60%, both of which tend to be stable.

[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for simulating the casting of molten aluminum alloy, characterized in that, include: Step 1: Solve the fundamental physical fields of the casting, filling, and solidification processes, outputting the geometric position of the free surface, the velocity and temperature of the melt at the free surface, and the velocity field and solidity field of the bulk melt step by step. Step 2: Discretize the free surface into surface material entities corresponding to a single membrane. Construct a dynamic surface material entity graph using each surface material entity as a node and the connectivity between adjacent membranes as edges. Each node carries the membrane's position, unit normal vector, cumulative exposure time since exposure to the atmosphere, and features indicating the wetting state of the two membrane surfaces. Create new surface material entities at newly generated free surfaces, ensuring the individual identity and folding topology of the membranes continue across time steps. Step 3: Input the dynamic surface material entity graph into a graph neural network's topological event recognition network, and then... The message transmission outputs the topological event types of free surface self-folding and bonding, liquid front collision and merging, and injection fracture and splashing, as well as the diaphragm pairs participating in the events; Step four, reconstruct the network with the diaphragm pairs participating in each topological event and their contact geometry input configuration states, and output the configuration state of the formed entangled diaphragm as a double-layer film or a single-layer film, whether the two film surfaces of the double-layer film are welded, its spatial orientation relative to the casting coordinate system, and its maturity; Step five, take the entangled diaphragm with the configuration state as the film entity inside the bulk phase, and make it Lagrangian advance with the velocity field of the bulk melt, and freeze it together with the configuration state when the solid fraction at its location reaches a threshold; Step six, statistically analyze the configuration states of the film entities frozen at various locations in the casting, generate a harmful double-layer film distribution field, and combine it with the stress field to predict casting defects.

2. The aluminum alloy melt casting simulation method according to claim 1, characterized in that, In step one, while solving the fundamental physical field, the surface age transport equation is also solved on the free surface, so that each point on the free surface carries the cumulative exposure time since it was exposed to the atmosphere and generated. The source term of the surface age transport equation determines the film thickness growth rate based on the local temperature and oxygen partial pressure of the free surface according to the oxidation growth kinetics, so that the exposure time corresponds to the thickness and strength of the oxide film.

3. The aluminum alloy melt casting simulation method according to claim 1, characterized in that, In step two, each surface material entity undergoes a Lagrangian progression between time steps with the melt velocity at the free surface, and its area is updated according to the local stretching rate of the free surface, thereby superimposing the Lagrangian description of the free surface on the Eulerian description of the bulk melt; the features carried by the node also include the area of ​​the film, the local temperature, and the wetting state indicating that its two film surfaces face inward toward the inside of the melt and outward toward the air gap, respectively. The edges connecting the nodes corresponding to adjacent films in the dynamic surface material entity diagram also carry the relative orientation between the adjacent films.

4. The aluminum alloy melt casting simulation method according to claim 1, characterized in that, In step three, the topology event recognition network is trained in a supervised manner on a sample set that has been generated by solving the basic physical field and tracking the high-resolution interface and has been labeled with the topology event type and the membrane pairing relationship. For membrane pairs that are close to each other, when their unit normal vectors are opposite in direction and the distance is less than a set threshold, they are determined to be candidate events of the free surface self-folding and bonding or the liquid front collision and merging, and are handed over to the topology event recognition network for further classification.

5. The aluminum alloy melt casting simulation method according to claim 1, characterized in that, In step four, the configuration state reconstruction network determines whether the formed membrane is a double-layer unwelded air-clamped folded state based on whether the membrane surfaces of the two membranes involved in the topological event are both non-wetting membrane surfaces, whether an air gap is sandwiched in the middle during the bonding, and the strength and weldability of the membrane determined by the exposure time of each of the two membranes; it determines the spatial orientation of the formed membrane relative to the casting coordinate system based on the geometric orientation of the bonding surfaces of the two membranes at the time of the event; and it determines the maturity of the formed membrane based on the larger of the exposure time of the two membranes.

6. The aluminum alloy melt casting simulation method according to claim 5, characterized in that, For the topological event of the liquid front collision and merging, the configuration state reconstruction network further outputs whether the two oxide films are welded when they meet and the configuration of the oxide joint formed when they are not welded, so as to characterize the unwelded oxide joint formed when the thin-walled part or multi-gate part is filled.

7. The aluminum alloy melt casting simulation method according to claim 1, characterized in that, In step five, as the membrane entity advances with the velocity field of the bulk melt, it updates its configuration state based on the contact events it subsequently experiences. For membrane entities that subsequently undergo bonding and stacking, their bilayerity and spatial orientation are updated according to steps three and four to reproduce the secondary folding of the membrane within the bulk phase.

8. The aluminum alloy melt casting simulation method according to claim 1, characterized in that, In step six, the harmful bilayer film distribution field outputs the density, spatial orientation, and maturity of the harmful bilayer film at each spatial location; the harmful bilayer film distribution field is combined with the stress field and solidity field at that spatial location to predict the tendency for thermal cracking, porosity, and fatigue crack initiation at that spatial location, wherein the tendency for fatigue crack initiation is determined based on the density of the harmful bilayer film, the angle between its spatial orientation and the principal stress direction at that spatial location, and its maturity.

9. The aluminum alloy melt casting simulation method according to claim 1, characterized in that, The process also includes step seven, in which the pouring temperature, holding time, gate position and filling speed are the variables to be optimized, and the distribution of harmful double film and casting defect tendency predicted in step six are the optimization targets. The variables to be optimized are iteratively adjusted in the process parameter space and steps one to six are repeated until the distribution of harmful double film and casting defect tendency meet the set requirements, and the optimized process parameters are output.

10. A simulation system for molten aluminum alloy casting, characterized in that, include: The fundamental physics field solution module is used to solve the fundamental physics field of the casting and solidification process. It outputs the geometric position of the free surface, the velocity and temperature of the melt at the free surface, and the velocity field and solidity field of the bulk melt step by step. The surface material entity graph construction module is used to discretize the free surface into surface material entities corresponding to a membrane. It constructs a dynamic surface material entity graph with each surface material entity as a node and the connection relationship between adjacent membranes as an edge. The node carries the position of the membrane, the unit normal vector, the cumulative exposure time since exposure to the atmosphere, and the features that identify the wetting state of the two membrane surfaces. It also creates new surface material entities at the newly generated free surface, so that the individual identity and folding topology of the membrane are continued across time steps. The topology event recognition module is used to input the dynamic surface material entity map into a graph neural network for topology event recognition. After message passing, it outputs the topology event types such as free surface self-folding and bonding, liquid front collision and merging, and injection fracture and splashing, as well as the diaphragm pairs participating in the events. The configuration state reconstruction module is used to input the diaphragm pairs participating in each topology event and their contact geometry into the configuration state reconstruction network. It outputs the configuration state of the formed rolled-in diaphragm as a double-layer or single-layer diaphragm, whether the two surfaces of the double-layer diaphragm are welded, its spatial orientation relative to the casting coordinate system, and its maturity. The diaphragm tracking and freezing module is used to treat the rolled-in diaphragm with the configuration state as an internal diaphragm entity in the bulk phase. It is Lagrangian-propelled with the velocity field of the bulk melt and frozen along with the configuration state when the solidity at its location reaches a threshold. The defect prediction module is used to statistically analyze the configuration states of the diaphragm entities frozen at various locations in the casting, generate a harmful double-layer diaphragm distribution field, and predict casting defects in combination with the stress field.

Images