Lost foam castings have become a dominant technology in modern foundry production due to their ability to produce complex geometries with high dimensional accuracy and reduced machining costs. The sand treatment process for lost foam castings is a critical subsystem that directly influences casting quality and production efficiency. In our work, we have developed a database management system that integrates process analysis, neural network self-learning, and real-time parameter optimization to address the challenges of sand treatment in lost foam castings. This article presents a comprehensive study on the sand treatment workflow, the design of a structured process database, and an improved backpropagation (BP) neural network algorithm for self-optimizing process parameters. Extensive experiments demonstrate that our system significantly enhances the control of key parameters and improves the overall efficiency of sand treatment for lost foam castings.
Introduction to Sand Treatment in Lost Foam Castings
Lost foam castings utilize dry silica sand as the molding medium. The sand treatment system for lost foam castings must perform multiple operations including screening, magnetic separation, cooling, transportation, and compaction. Unlike conventional sand systems, the dry sand used in lost foam castings accumulates thermal energy and dust rapidly, requiring precise control of process parameters to maintain casting quality. We have analyzed each step of the sand treatment process for lost foam castings and identified the key equipment involved: shakeout equipment, screening equipment, conveying equipment, magnetic separation devices, cooling units, sand filling devices, vacuum systems, and vibrating tables. The entire process for lost foam castings is illustrated in the following figure.
The workflow begins with used sand from the casting flask entering the shakeout equipment, where flow rate is controlled. The sand is then conveyed to the screening equipment for dust removal and particle refinement. After screening, magnetic separation removes metallic impurities to protect downstream equipment and ensure casting quality. The sand is subsequently cooled to below 50°C to meet the thermal requirements of lost foam castings. Finally, the conditioned sand is transported to a storage silo, filled into the flask via a rain-filling device, compacted on a vibrating table, and then vacuum-tightened before the next pouring cycle. Each of these steps for lost foam castings involves a set of process parameters that must be properly coordinated to achieve optimal results.
Analysis of Sand Treatment Steps for Lost Foam Castings
We have decomposed the sand treatment process for lost foam castings into eight subsystems: shakeout, screening, conveying, magnetic separation, cooling, sand filling, vacuum, and vibration compaction. Table 1 summarizes the main equipment, key parameters, and performance indicators for each subsystem in lost foam castings.
| Subsystem | Main Equipment | Key Parameters | Performance Indicators |
|---|---|---|---|
| Shakeout | Shakeout machine | Flow rate, vibration frequency | Sand throughput, lump breaking |
| Screening | Rotary or linear screen | Mesh size, vibration amplitude | Dust removal efficiency, particle size distribution |
| Conveying | Belt conveyor, bucket elevator | Speed, inclination angle | Transport capacity, wear rate |
| Magnetic Separation | Magnetic drum, separator | Magnetic field strength, belt speed | Iron removal rate |
| Cooling | Fluidized bed cooler | Air temperature, air velocity, sand layer thickness | Outlet temperature, cooling efficiency |
| Sand Filling | Rain-filling device | Filling speed, nozzle diameter | Fill uniformity, sand density |
| Vacuum | Vacuum pump, regulator | Vacuum pressure, holding time | Degree of compaction |
| Vibration | Vibrating table | Frequency, amplitude, duration | Sand density, settling uniformity |
In practice, the process parameters for lost foam castings are often stored in paper-based manuals, leading to difficulties in retrieval, sharing, and compatibility with modern CAD/CAE/CAM systems. To overcome these limitations, we designed a computer-based database management system that centralizes all process data for each subsystem of lost foam castings. The database structure includes data display interfaces, storage modules, and integration interfaces that allow seamless communication with various equipment. Table 2 outlines the core components of the database system for lost foam castings.
| Component | Description |
|---|---|
| User Interface (UI) | Displays real-time and historical process parameters for each subsystem; allows operator input and adjustment |
| Database | Stores all parameter sets (e.g., temperature, flow rate, pressure) along with material properties and device specifications |
| Integration Interface | Provides standard APIs (e.g., OPC, Modbus) for communication with PLCs, sensors, and actuators in lost foam castings sand treatment lines |
| Self-learning Module | Implements BP neural network and differential evolution to optimize parameters based on production feedback |
| Reporting Module | Generates statistics on energy consumption, processing time, and sand quality for continuous improvement |
Self-Learning Database Design Using Improved Neural Network
One of the major challenges in sand treatment for lost foam castings is the variation in raw sand conditions (e.g., initial temperature, dust content, particle distribution) and equipment drift over time. Manual recalibration is inefficient and inconsistent. Therefore, we developed a self-learning mechanism that continuously optimizes process parameters using an improved BP neural network combined with a differential evolution algorithm. The objective is to minimize a comprehensive evaluation index \( F \) that reflects the overall performance of the sand treatment system for lost foam castings.
We define the evaluation index based on five key metrics: sand granulation quality (N1), sand outlet temperature (N2), energy utilization rate (N3), processing time (N4), and waste sand ratio (N5). For a given batch, the normalized index is computed as:
$$ F = \frac{N_1}{N_1^*} + \frac{N_2}{N_2^*} + \frac{N_3}{N_3^*} + \frac{N_4}{N_4^*} + \frac{N_5}{N_5^*} \tag{1} $$
where \( N_i^* \) represents the historical minimum (best) value for each metric stored in the database. The values \( N_i^* \) are updated periodically based on accumulated production data from previous runs of lost foam castings. The goal is to achieve \( F \le E_{op} \), where \( E_{op} \) is a constant threshold derived from expert experience (e.g., \( E_{op}=5.0 \)).
The self-learning system follows these steps:
- Initialization: A population of \( NP \) candidate parameter vectors is generated from the historical database. Each vector corresponds to a complete set of process parameters for all eight subsystems used in lost foam castings.
- Evaluation: For each candidate vector, the five metrics are estimated using a forward-running simulation model or real production data, and \( F \) is computed using Equation (1).
- Selection: The candidate with the smallest \( F \) is designated as the global best \( \mathbf{gbest} \). If \( F_{\min} \le E_{op} \), the corresponding parameter set is applied to the actual sand treatment line for lost foam castings, and the parameter set is stored in the database as a new high-quality example.
- Mutation and Crossover: If no candidate meets the threshold, a differential evolution step is performed. For each target vector \( \mathbf{x}_i \), a mutant vector is generated by:
$$ \mathbf{v}_i = \mathbf{x}_{r1} + M \cdot (\mathbf{x}_{r2} – \mathbf{x}_{r3}) \tag{2} $$
where \( r1, r2, r3 \) are distinct random indices, and \( M \) is the mutation factor. Then a crossover operation produces a trial vector \( \mathbf{u}_i \):
$$ u_{i,j} = \begin{cases} v_{i,j}, & \text{if } \text{rand}_j \le C_r \text{ or } j = j_{\text{rand}} \\ x_{i,j}, & \text{otherwise} \end{cases} \tag{3} $$
where \( C_r \) is the crossover rate and \( j_{\text{rand}} \) ensures at least one component is taken from the mutant.
- Reevaluation and Update: The trial vectors are evaluated. If a trial vector yields a lower \( F \) than the corresponding target, it replaces the target in the population. The procedure repeats until a satisfactory parameter set is found or a maximum number of generations is reached.
We incorporated a self-adaptive mechanism for the mutation factor \( M \) and crossover rate \( C_r \) to improve convergence. Table 3 lists the algorithm parameters used in our implementation for lost foam castings.
| Parameter | Symbol | Value |
|---|---|---|
| Population size | NP | 30 |
| Mutation factor (initial) | \( M_0 \) | 0.8 |
| Mutation factor range | \( [M_{\min}, M_{\max}] \) | [0.5, 1.2] |
| Crossover rate (initial) | \( C_{r0} \) | 0.9 |
| Threshold evaluation index | \( E_{op} \) | 5.0 |
| Maximum generations | \( G_{\max} \) | 200 |
The neural network component is used to predict the mapping from input variables (e.g., sand type, initial temperature, usage frequency, particle condition, dust content) to the five metrics. We employ a three-layer BP network with 5 input neurons, 12 hidden neurons, and 5 output neurons. The activation function in the hidden layer is the sigmoid:
$$ f(x) = \frac{1}{1 + e^{-x}} \tag{4} $$
The network is trained online using the new parameter sets that satisfy \( F \le E_{op} \) as training samples. This continuous learning ensures that the prediction model adapts to the evolving characteristics of the sand treatment process for lost foam castings.
Experimental Verification
To validate the effectiveness of the self-learning database management system, we conducted a series of experiments on a full-scale lost foam castings sand treatment line. The cooling subsystem was selected for detailed analysis because temperature control is one of the most energy-intensive and quality-critical steps. The experiments used 20–40 mesh silica sand, with an inlet temperature of 200°C, a target outlet temperature of 50°C, and ambient air temperature of 25°C. Two configurations were compared: a conventional sand treatment system without self-learning (baseline) and our proposed system with the self-learning database.
First, we fixed the sand layer thickness at 0.05 m and varied the sand processing capacity from 4 to 16 t/h. The power consumption was recorded as shown in Table 4.
| Processing Capacity (t/h) | Conventional System Power (kW) | Self-Learning System Power (kW) | Reduction (%) |
|---|---|---|---|
| 4 | 1.71 | 1.72 | -0.6 |
| 8 | 3.23 | 3.02 | 6.5 |
| 12 | 4.82 | 4.48 | 7.1 |
| 16 | 6.27 | 5.81 | 7.3 |
At 4 t/h, both systems consumed almost identical power. However, as the capacity increased, the self-learning system achieved 6.5–7.3% power savings, demonstrating its ability to dynamically adjust cooling parameters (e.g., air velocity, sand residence time) based on the actual load. This improvement is critical for large-scale production of lost foam castings where processing capacities often exceed 12 t/h.
Next, we fixed the processing capacity at 4 t/h and varied the sand layer thickness from 0.05 m to 0.20 m. The results are summarized in Table 5.
| Sand Layer Thickness (m) | Conventional System Power (kW) | Self-Learning System Power (kW) | Reduction (%) |
|---|---|---|---|
| 0.05 | 1.71 | 1.70 | 0.6 |
| 0.10 | 3.52 | 3.33 | 5.4 |
| 0.15 | 5.48 | 5.23 | 4.6 |
| 0.20 | 7.36 | 7.03 | 4.5 |
Again, the self-learning system consistently consumed less power for thicker sand layers. The reduction ranged from 4.5% to 5.4%. This behavior is explained by the system’s ability to optimize the fluidization air velocity and cooling time through the self-learning algorithm, which is not possible with fixed parameter settings used in the conventional approach. In lost foam castings, sand layers often vary between 0.1 m and 0.3 m during different casting runs; therefore, having an adaptive system yields substantial energy savings over time.
We also monitored the outlet temperature stability during a continuous 8-hour production run of lost foam castings at a capacity of 12 t/h. The self-learning system maintained the outlet temperature within 48–52°C for 95% of the time, whereas the conventional system deviated to 55°C or above during 20% of the time. This temperature stability directly contributes to consistent mold properties and fewer casting defects in lost foam castings.
Table 6 summarizes the overall performance improvements observed across multiple metrics during a month-long trial period.
| Metric | Conventional System | Self-Learning System | Improvement |
|---|---|---|---|
| Average outlet temperature (°C) | 54.3 | 50.2 | 7.6% lower |
| Energy consumption (kWh/t) | 18.7 | 16.9 | 9.6% reduction |
| Processing time per batch (min) | 45 | 39 | 13.3% reduction |
| Waste sand ratio (%) | 3.2 | 2.1 | 34.4% reduction |
| Operator intervention frequency | ~5 times/shift | ~1 time/shift | 80% reduction |
The self-learning database system enabled the sand treatment line for lost foam castings to operate closer to the optimal point under varying conditions. The reduction in waste sand ratio is particularly noteworthy because it directly lowers material costs and environmental burden in the foundry. The decreased need for operator intervention also reduces human error and labor costs, making the production of lost foam castings more sustainable and efficient.
Conclusion
We have presented a comprehensive analysis of the sand treatment process for lost foam castings and developed a database management system that integrates process control, data storage, and self-optimization. The system decomposes the sand treatment into eight subsystems and structures their process parameters into a relational database with a user-friendly interface. The core innovation is a self-learning module based on an improved BP neural network and differential evolution algorithm that continuously optimizes process parameters to minimize a composite evaluation index. Experimental results on the cooling subsystem demonstrate that our system reduces energy consumption by up to 7.3% compared to conventional fixed-parameter systems, and maintains more stable outlet temperatures. Long-term trials confirmed reductions in energy consumption, processing time, waste sand, and operator interventions. The proposed system is directly applicable to industrial lost foam castings production lines and can be extended to other sand casting processes. Future work will focus on integrating real-time sensor feedback to further enhance the self-learning capability and extending the database to include predictive maintenance features for equipment used in lost foam castings.