Flexible Behavior of Transverse Joints in Full-Scale Precast Concrete Slabs with Open-Type Joint Method

Abstract

Cracks and concentrated stresses can develop in precast concrete slabs, depending on the quality of the joints. The open-type joint method was adopted herein to fabricate a full-scale precast concrete slab joint. The open-type joint method features an exposed joint configuration that allows for direct installation of shear connectors without temporary formwork, improving constructability and load transfer efficiency. Full-scale load testing was carried out using a four-point loading experiment, revealing that the precast concrete slab had a yield load of 550 kN and maximum load of 733 kN. A slab using the cast-in-place method was measured to have a yield load of 500 kN and maximum load of 710 kN. A finite element analysis (FEA) model modeled the precast concrete slab, and the displacement and maximum load were analyzed. The FEA showed a maximum error within 7%. Therefore, the FEA results can predict the structural performance of the load–displacement of the precast concrete slab. The support vector regression model predicted key structural performance indicators such as concrete compressive strength, maximum load, displacement, and principal stress. The prediction results indicated that the average error converged within 3%. The prediction results of the SVR model can complement FEA by estimating outcomes without the need for complex modeling. Thus, the precast concrete slab using the open-type joint method was able to achieve structural performance equivalent to that of the slab using the cast-in-place technique. Furthermore, FEA and machine learning will be able to predict the structural performance of precast concrete slabs using the open-type joint method.

1. Introduction

The precast concrete (PC) technique entails the assembly and installation of prefabricated components at the construction site. The PC technique enhances construction efficiency, guarantees consistent quality, and mitigates the effects of weather conditions [1,2]. Consequently, the PC method is extensively utilized in diverse structures, including bridges, buildings, and industrial facilities [3,4,5,6]. In particular, slab forms are easy to manufacture and transport. Moreover, PC slabs can reduce construction time and lower costs, thereby ensuring both the economic feasibility and constructability of the structure [7,8,9]. Accordingly, the PC method is being proposed as an alternative that can replace the traditional cast-in-place (CIP) method [10,11,12,13,14,15,16].

On the other hand, the CIP method has been used for a long time by directly pouring concrete on-site. However, the CIP method has limitations such as long construction time and difficulty in ensuring uniform quality [17,18,19,20,21,22]. Park et al. [23] pointed out that the unevenness of the mix occurring in the CIP method decreases strength and durability, which can result in long-term structural problems. Furthermore, Trentini et al. [24] reported that differential shrinkage and cracking occurring during the construction process hinder the durability of the structure. Due to these issues, the necessity of utilizing the PC method is being emphasized even more.

However, the structural performance of PC slab joints has recently emerged as an important issue during the assembly process of PC [25,26,27,28,29]. Especially if the stiffness and integrity of the joint are not sufficiently secured, cracks may occur and load transfer may fail, leading to a decrease in safety. Accordingly, research to improve the structural performance of PC slab joints is continuously being conducted [30,31,32,33,34,35,36,37,38,39,40]. Traditional precast concrete slabs (PS) used the pocket-type joint method to place rebar at uneven intervals. However, this method does not evenly distribute stress and concentrates it in specific areas, increasing the likelihood of crack formation. Additionally, construction precision within the shear pocket is required, and if securing the rebar anchorage length is difficult, a decrease in strength may occur. To address these issues, the open-type joint method has been proposed. This method applies an open-type shape to the joint of the PS, allowing the rebar to be uniformly spaced and thereby evenly distributing stress. Abdullah and Mosheer [41] conducted experiments on various rebar placements. The experimental results indicated that uniformly spaced rebar placement significantly improved the shear performance of the PS. Hillebrand et al. [42] also reported that even when long-term loads are applied to this type of PS joint, the concrete slab effectively distributes the load. Although similar joint systems have been introduced internationally, the open-type joint method presented in this study is newly proposed within the domestic context and has not been previously validated through full-scale structural testing.

However, full-scale large structure experiments are expensive, take a long time to complete, and have limitations when it comes to analyzing different variables. Consequently, FEA was used in this study to perform a nonlinear analysis. The FEA technique has been used in many studies to comprehend and forecast structural behavior [43,44,45,46]. Nonlinear analysis is especially important for assessing the nonlinear behavior of rebar and concrete, and it is also vital for forecasting how well real structures will perform. Hence, factors like concrete cracking, rebar yielding, and the interaction between concrete and rebar should be taken into account in the nonlinear analysis of concrete structures [47]. A nonlinear finite element model of concrete with nonlinear behavior was proposed by Xue et al. [48]. This model describes how nonlinear analysis has a substantial influence on the long-term performance prediction of real structures while also explaining the fracture mechanism and deformation laws of concrete. In a nonlinear analysis of reinforced concrete (RC) beams, Park et al. [23] contrasted the actual crack patterns with the predicted load–displacement responses. Their study showed that the behavior of real structures can be accurately replicated using nonlinear analysis. In the present research, a three-dimensional (3D) model of a PS was constructed using the open-type joint method through FEA in order to analyze the maximum load and displacement.

Recently, studies on predicting the strength and behavior of concrete structures using machine learning (ML) have been recently conducted [49,50,51]. Zhang et al. [52] conducted a study using an extreme learning machine (ELM) to predict the strength of rubber fiber concrete. The study results demonstrated that ELM has a fast modeling speed and high prediction accuracy. Ahmad et al. [53] used an artificial neural network (ANN) to predict crack load, bending strength, and shear strength. Experimental results indicated that the ANN produced results consistent with those obtained from traditional models. This study demonstrated that ML could contribute to the design and safety enhancement of structures. Chaabene et al. [54] reviewed four models—ANN, support vector machine (SVM), decision tree, and evolutionary algorithm (EA) to predict the mechanical strength of concrete. The results suggested that ML has the potential to complement the limitations of traditional prediction methods. However, it was argued that in the field of civil engineering, it is difficult to fully utilize the functions of ML; thus, continuous research is needed. The present study used ML to predict the finite element analysis results of PS. Among the ML models, support vector regression (SVR) can be trained with a small amount of data and has high prediction accuracy. Therefore, this research was conducted using SVR. Compared to ANN and random forest, SVR offers better generalization with limited data and continuous output prediction, making it suitable for structural response modeling. The training data used were compressive strength, Young’s modulus, yield stress, cracking strain, displacement, maximum load, and stress derived from ABAQUS (2021) [55]. Finally, the SVR model was used to predict the displacement, maximum load, and stress of the PS joint based on its compressive strength, and the predicted material properties were compared with actual experimental and finite element analysis results to verify the model’s reliability.

According to previous studies, the open-type joint method can enhance the shear performance and load transfer capacity of concrete slabs. However, there is a lack of research that quantitatively analyzes the structural performance under various load conditions. Therefore, this study fabricated full-scale PS using the open-type joint method and slabs using the CIP technique and evaluated their structural performance through four-point loading tests. The PS was studied in comparison to the slab using the CIP method to examine its structural behavior. In addition, the PS using the open-type joint method was constructed as a three-dimensional (3D) nonlinear FEA model, and the reliability of the analysis model was verified by comparing it with experimental results. Finally, an SVR model was constructed based on the FEA analysis data to predict the key finite element analysis results of the PS joint. The applicability of SVR was examined in the design process of PC structures using the open-type joint method.

2. Materials and Methods

2.1. Fabrication

Figure 1 illustrates a PS utilizing the open-type joint method. The joint between the slabs is configured in an open-type design, facilitating the placement of shear connectors at fixed intervals. This configuration is advantageous in mitigating stress concentration and maintaining structural integrity. It improves structural integrity by enabling uniform rebar anchorage across the joint, which facilitates consistent shear transfer, minimizes abrupt stiffness changes, and effectively reduces localized stress concentrations. The underside of the slab is uniformly reinforced with rebar, ensuring a distinct load transfer pathway and enhancing shear performance and durability. This study fabricated a full-scale slab employing the open-type joint method. The structural performance of the PS employing the open-type joint method was analyzed by comparison with a slab utilizing the CIP method.

Figure 2 shows the slab side and upper road with dimensions of 5660 × 1970 × 240 mm, and the joint width is set to 600 mm. The specimen was designed based on structural calculations performed according to the concrete design standards and concrete bridge design standards [56,57]. The reinforcement bars used were H19, and the yield strength of the reinforcement bars was 400 MPa. The concrete used in the joint was non-shrinkage concrete with 27 MPa and 40 MPa, respectively, for each parameter.

Figure 3 shows the process of fabricating the specimen. First, rebars were placed in the fabricated mold, and a strain gauge was installed (Figure 3a). Figure 3b shows the specimen being cured by pouring 27 MPa concrete into the mold. After the concrete curing was completed, the formwork was removed, and the specimen was demolded (Figure 3c). Finally, the completed specimens were inspected for any defects before the experiment (Figure 3d).

The four-point loading test analyzed the bending performance of the transverse joint between the slab using the CIP method and the slab using the PC method. Figure 4 shows the setup of the four-point loading experiment. The actuator’s capacity was 2000 kN, and its total length was 3214 mm. The total length from the end of the actuator to the bottom of the specimen fixture was set to 6300 mm. A steel rod with a diameter of 40 mm and a length of 600 mm was installed between the actuator and the specimen. In addition, a loading plate measuring 150 × 19 × 600 mm was placed at the bottom of the steel rod to apply the load. The load was applied longitudinally using an actuator to induce stress throughout each specimen. The loading method was applied at four points, considering the tire ground pressure based on the vehicle load [58]. The actuator applied the load in displacement-controlled mode at a rate of 1 mm/min [59]. The load application points were applied at 600 mm intervals from the center of the top of the specimen (Figure 4a). The slab specimen was supported in a simply supported condition using steel frames and rubber pads at both ends, allowing vertical displacement and rotation while minimizing edge restraint. Experimental data were collected using a data logger (TDS-540), and the crack modes of the specimen were analyzed using the digital image correlation (DIC), technique. A spray was applied to the surface of the specimen to form a speckle pattern, and a high-speed PCIe camera was used to capture images at a rate of 1 frame per second to measure strain and displacement (Figure 4b). The specimens were categorized according to the method and strength used, as shown in

Table 1. Test parameters.

Test Specimen	Specifies	Compressive Strength of Concrete
CS-27 (test)	Cast-in-place slab (CS)	27 MPa
PS-27 (test)	Precast concrete slab (PS)	27 MPa
PS-40 (test)	Precast concrete slab (PS)	40 MPa

. Except for the joint regions, the rest of the slab was fabricated using concrete with a compressive strength of 27 MPa. For example, the slab using the CIP method with a strength of 27 MPa was designated as “CS-27,” and the slab using the precast method with a strength of 40 MPa was designated “PS-40.” Strain gauges were categorized by the slab type and installation location; the details are provided in

Table 2. Status of gauge installation.

Gauge Name	Slab Types	Gauge Locations	Gauge Locations
CS-T-27	Cast in place (CS)	Strain gauge top (T)	27 MPa
CS-B-27		Strain gauge bottom (B)
PS-T-27	Precast concrete slab (PS)	Strain gauge top (T)	27 MPa
PS-T-40			40 MPa
PS-B-27	Precast concrete slab (PS)	Strain gauge bottom (B)	27 MPa
PS-B-40			40 MPa

. Strain gauges and linear variable displacement transducers (LVDTs) were installed on each specimen to measure the deformation and displacement of the specimens. The strain gauges were attached to the top and bottom of both CS and PS, while the LVDT was installed at the center bottom of the slab. All strain gauges were calibrated prior to installation according to manufacturer loading test analyzed. The measurement error margin was within ±1%, as stated by the strain gauge manufacturer.

2.2. Experimental Result

Figure 5 shows the displacement graph of the specimen according to the load. All three specimens exhibited similar initial stiffness. Additionally, as the load increased, nonlinear behavior after yielding was observed. The yield load was measured at 500 kN for CS-27 and 550 kN for PS-40. The maximum load of CS-27 was 710 kN, and the maximum load of PS-40 was approximately 730 kN. PS-27 had a yield load measured at 480 kN and a maximum load of 600 kN, showing the lowest structural performance among the three specimens. PS-40 exhibited similar structural behavior when compared to CS-27. This behavior can be attributed to the improved material properties of the 40 MPa non-shrinkage concrete used in PS-40, which contributed to increased load-carrying capacity and enhanced bonding performance at the joint. In contrast, PS-27 exhibited lower performance due to the relatively low strength of the concrete, which may have resulted in earlier cracking and reduced joint integrity. These differences highlight the critical role of concrete compressive strength in ensuring the structural effectiveness of precast slab joints. As a result, when the strength of non-shrinkage concrete is above 40 MPa, the PS joint can achieve structural performance similar to that of a slab joint constructed using the CIP method.

Figure 6a compares the strain measured at the lower central part of the specimen according to the load. The strain of CS-B-27 (Test) occurred from a load of 100 kN, and ultimately a large tensile deformation of about 15,000 με was observed. PS-B-27 (Test) showed deformation starting from approximately 140 kN, with a maximum strain measured at about 7800 με. The maximum strain of PS-B-40 (Test) was approximately 4000 με. At the yield load of 480 kN for PS-B-27 (Test), the strain for PS-B-40 (Test) was the lowest at 1100 με. The strain of PS-B-27 (Test) was measured at 2000 με, while the strain of CS-B-27 (Test) was measured at 3000 με, the highest among the tests. Therefore, CS-B-27 (Test) was expected to be disadvantageous in terms of structural stability due to excessive strain increase. PS-B-40 (Test) exhibited the lowest maximum strain, indicating superior deformation resistance under load. Hence, the open-type joint PS is thought to have better structural performance in terms of rebar strain than the slab made with the CIP method.

Figure 6b compares the strain measured at the top of the joint of the specimen according to the load. Overall, the strain was measured to be a maximum of about 25 με or less in all three specimens. When the load was applied to CS-27, a strain level of approximately 10–15 με occurred. PS-27 showed approximately 20 με at loads over 200 kN. PS-40 also showed the highest strain level of 25 με among the specimens under loads exceeding 200 kN. Overall, all three specimens exhibited minimal strain at the top of the joint. These results are due to the joint exhibiting structurally integral behavior. PS-40 experienced a somewhat larger strain compared to CS-27, but the difference was limited to within 5 με. Therefore, PS-40, which applied the open-type joint method, can achieve a joint performance level similar to that of CS-27.

Figure 7a illustrates the crack distribution and failure modes of CS-27 as the load increases. The crack occurrence pattern of the CS-27 specimen showed minimal cracks at the initial load stage of 140 kN. At a load of 300 kN, vertical cracks appeared in the center of the slab. As the load increased to 500 kN, the cracks gradually expanded, and shear cracks began to appear, especially near the supports. A sudden diagonal crack occurred at the node at the maximum load of 712 kN, leading to the failure. Furthermore, the strain on the rebar remained at a minimal level in the initial stage, but as the load increased, it tended to concentrate locally in the center of the slab. Failure occurred at the maximum load due to a diagonal crack caused by the increased strain at the node.

Figure 7b illustrates the crack distribution and failure modes of PS-27 as the load increases. The crack propagation pattern of the PS-2,7 specimen did not show any distinct cracks at the initial load stage of 100 kN. Under a load of 300 kN, vertical cracks appeared in the center of the slab. At 500 kN, the existing cracks further expanded, and new web shear cracks formed near the supports. Rapid shear cracks occurred at the node at the maximum load of 596 kN, and failure progressed as the shear strength of the concrete reached its limit. The strain was generally low and uniformly distributed in the initial stage. As the load increased, the strain on the rebar in the center of the slab increased, and deformation between the rebar and concrete progressed due to the bending moment. Therefore, PS-27 initially developed cracks in the central part of the slab due to the bending moment, but as the load increased, shear cracks were activated and shear failure became dominant.

Figure 7c illustrates the crack distribution and failure modes of PS-40 as the load increases. At the initial load stage of 145 kN, no clear cracks were formed in the specimen. At a load level of 300 kN, the first vertical crack was observed in the center of the slab. At 500 kN load, a new diagonal crack appeared near the support. At the maximum load of 733 kN, the depth of the crack gradually increased as the load increased. The strain of the rebar was uniformly distributed during the initial loading stage and maintained a generally low value. But after 300 kN, the strain concentrated in the slab’s center due to the rebar’s increased strain from the bending moment. At 500 kN load, the strain of the rebar in the center of the slab increased, but ultimately, the strain was evenly distributed throughout the entire slab.

Consequently, because PS-40 exhibited fewer cracks than CS-27 in the final failure mode, the slab using the open-type joint method has superior load distribution capability compared to the slab made with the CIP method. Furthermore, while PS-40 showed a uniform distribution of strain overall, CS-27 had localized strain concentration in the center of the slab, indicating that the structural performance of the PS using the open-type joint method is relatively superior.

3. Finite Element Analysis

3.1. 3D Modeling

FEA was performed to analyze the structural behavior of PS with the open-type joint method. Before the analysis, the parameters for the comparative analysis were listed as shown in

Table 3. Test parameters (FEA).

Test Specimen	Specifies	Compressive Strength
CS-27 (test)	Cast in place (CS)	27 MPa
PS-40 (test)	Precast concrete slab (PS)	40 MPa
PS-40 (FEA)	Precast concrete slab (PS)	40 MPa

. The analysis utilized the general-purpose structural analysis software ABAQUS (2021). The total size of the analysis model was set to 5660 × 1970 × 240 mm, the same as the experimental specimen. The modeled components consisted of a slab, joint, girder upper plate, and rebar. The slab, joint, and girder upper plate were implemented as C3D8R solid elements. The contact surfaces of each member cannot be considered fully integrated, so they were modeled using spring elements (connecting two nodes). Rebars were modeled using the T3D2 truss element, and the steel bars used for load application were modeled using the B31 beam element (Figure 8a). The support conditions for the underside of the slab were set to the ground-spring type, identical to the experiment. To realistically reflect the ground behavior, nonlinear springs were applied with stiffness values derived from assumed soil conditions. The mesh of the analysis model was set to h/L = 0.1 for curvature control and a size of 100 mm. The mesh shape was applied with a structured hex mesh to prevent twisting (Figure 8b). The rebars were configured to be embedded within the concrete using the embedded element technique. The load was set to 800 kN and applied at 600 mm intervals (Figure 8c).

The material properties applied in the analysis are summarized in

Table 4. Material properties.

Steel
Elastic		Plastic
Young’ modulus	Poisson’s ratio	Yield stress	Plastic strain
200,000	0.300	275	0
		410	0.027
Concrete
Elastic
Young’ modulus		Poisson’s ratio
200,000		0.300

. The properties are based on the concrete damage plasticity (CDP) model [60,61,62,63,64]. The Young’s modulus used was 200,000 MPa, and Poisson’s ratio was 0.3. The yield stress was set to 275 and 410 MPa to reflect the range of plastic deformation. Additionally, the corresponding plastic strain values of 0 and 0.027 were used. The Young’s modulus of the concrete was applied as 36,800 MPa, and the Poisson’s ratio was 0.2. In addition, to reflect the nonlinear behavior of concrete, damage and inelastic strain under compression and tension states were applied according to

Table 5. Properties of 40 MPa concrete in compression and tension.

Compressive Behavior		Compression Damage
Yield Stress	Inelastic Strain	Damage Parameter	Inelastic Strain
20.4	0	0	0
25.6	0.00003	0	0.00003
30.0	0.00005	0	0.00005
33.6	0.00016	0	0.00016
36.4	0.00027	0	0.00027
38.4	0.00040	0	0.00040
39.6	0.00056	0	0.00056
40.0	0.00075	0	0.00075
39.6	0.00096	0.01	0.00096
38.4	0.00120	0.04	0.00120
36.4	0.00147	0.09	0.00147
33.6	0.00176	0.16	0.00176
30.0	0.00208	0.25	0.00208
25.6	0.00243	0.36	0.00243
20.4	0.00280	0.49	0.00280
14.4	0.00320	0.64	0.00320
7.6	0.00363	0.81	0.00363
Tensile behavior		Compression damage
Yield stress	Cracking strain	Yield stress	Cracking strain
4		0
0.04		0.0013333

. The compressive behavior was configured such that the damage parameter started at 0.01 and increased to 0.81 when the maximum compressive strength of 40 MPa was reached. The tensile behavior was configured to set the damage parameter to a maximum of 0.99, considering a cracking strain of 0.001333 at a tensile strength of 4 MPa.

3.2. FEA Result

Figure 9a shows the principal stress distribution for PS-40 (FEA). The principal stresses of PS-40 (FEA) are concentrated in the load application position, and tensile stresses are especially high at the bottom of the slab. These results indicate that bending behavior occurred in the central part of the slab under load, causing tensile forces to act on the lower part. Stress concentration also partially occurred at the joint and girder upper plate. As the load is applied to the center of the slab, tensile and compressive forces occur, and due to the reaction force, compressive forces are concentrated at the joint. Figure 9b shows the displacement distribution of PS-40 (FEA) under the same load conditions. The bending behavior under load application was concentrated in the center of the slab, resulting in maximum deflection. The displacement distribution showed a tendency to reach its maximum value at the center of the slab and gradually decrease toward both ends. In particular, the displacement near the girder upper plate was small. This is because the node condition effectively suppresses the load and vertical displacement of the slab. Figure 9c compares the central strain distribution of PS-40 measured using the DIC technique with the finite element analysis results obtained using ABAQUS. Both the experimental results and the analysis results showed a similar stress distribution with stress concentrated around the joint. The stress appeared relatively lower in PS-40 (FEA). This trend indicates that the nonlinear finite element analysis quantitatively reproduces the bending behavior observed in the experiments. Therefore, the proposed analysis model will be able to predict structural performance under various variable conditions.

Figure 10 is a load–displacement graph comparing the experimental results of CS-27 and PS-40 with the PS-40 (FEA) results. PS-40 (FEA) experienced yielding of the rebar at approximately 580 kN, and the maximum load of PS-40 (FEA) was 780 kN. The PS-40 (test) curve and the PS-40 (FEA) curve showed similar shapes in the nonlinear behavior after yielding, and there was a tendency for them to match at the point of maximum load and in the subsequent strength reduction phase. The FEA model showed slightly higher stiffness compared to the experiment, but the error in maximum load was within 6%, demonstrating the accuracy of the FEA model. Therefore, the FEA model proposed in this study can predict bending behavior.

4. Property Prediction Models

4.1. Support Vector Regression (SVR) Model

SVR was adopted as the machine learning model. The SVR model can effectively learn linear and nonlinear relationships through kernel functions and has high predictive performance even with a small amount of data. Hence, this study selected SVR to effectively utilize the limited training data. The learning structure of the SVR model was composed of two stages based on the nonlinear properties of concrete materials. Figure 11 shows the learning and evaluation procedure of the entire prediction model. The sequence was carried out as follows: data normalization, separation of training and validation data, hyperparameter initialization, model training and evaluation, model performance improvement, and final prediction. The first-stage model was designed to predict the compressive strength value using cracking strain, Young’s modulus, and yield stress as input variables. The second-stage model was designed to predict displacement, maximum load, and stress using the predicted compressive strength and FEA results.

$$\begin{array}{c}min{\displaystyle \frac{1}{2}}{‖w‖}^2+C{\sum }_{i=1}^n({\xi }_i+{\xi }_i^{\mathrm{*}})\\ \hspace{1em}\hspace{1em}\mathrm{Subject} \mathrm{to} \left\{\begin{array}{c}{y}_i-\left(w{x}_i+b\right)\le \epsilon +{\xi }_i\\ \left(w{x}_i+b\right)-y_i \le \epsilon +{\xi }_i\end{array}\right.\end{array}$$

(1)

Equation (1) is the optimization process of the SVR model. Here, w is the weight vector, C is the penalty coefficient for error tolerance, ε is the tolerance range, and ${\xi }_i$ and ${\xi }_i^{*}$ are slack variables that reflect the prediction error. After setting the hyperparameters ε and C of SVR, the prediction performance was improved through iterative training and evaluation.

$$\mathrm{M}\mathrm{S}\mathrm{E}={\displaystyle \frac{1}{n}}\sum _{i=1}^n{(y_i+{\widehat{y}}_i)}^2$$

(2)

Equation (2) is the calculation process of the mean squared error (MSE). Here, n is the number of data points, $y_i$ is the actual value, and ${\widehat{y}}_i$ is the predicted value. The prediction performance of the two-stage model was evaluated using MSE.

4.2. Model Training

The data used for model training are organized in

Table 6. Properties by compressive strength.

Compressive Strength	Young’s Modulus	Yield Stress	Cracking Strain
20	21,200	0.020	0.000943396
21	21,740	0.021	0.000965788
22	22,280	0.022	0.000988180
23	22,820	0.023	0.001010572
24	23,360	0.024	0.001032964
25	23,900	0.025	0.001055356
26	24,440	0.026	0.001077747
27	24,980	0.027	0.001100139
28	25,520	0.028	0.001122531
29	26,060	0.029	0.001144923
30	26,600	0.030	0.001167315
31	26,940	0.031	0.001183917
32	27,280	0.032	0.001200519
33	27,620	0.033	0.001217121
34	27,960	0.034	0.001233723
35	28,300	0.035	0.001250324
36	28,640	0.036	0.001266926
37	28,980	0.037	0.001283528
38	29,320	0.038	0.001300130
39	29,660	0.039	0.001316732
40	30,000	0.040	0.001333333
41	30,340	0.041	0.001349432
42	30,680	0.042	0.001365531
43	31,020	0.043	0.001381630
44	31,360	0.044	0.001397729
45	31,700	0.045	0.001413828
46	32,040	0.046	0.001429927
47	32,380	0.047	0.001446026
48	32,720	0.048	0.001462125
49	33,060	0.049	0.001478223
50	33,400	0.050	0.001494322
51	33,740	0.051	0.001509918
52	34,080	0.052	0.001525514
53	34,420	0.053	0.001541111
54	34,760	0.054	0.001556707
55	35,100	0.055	0.001572303
56	35,440	0.056	0.001587899
57	35,780	0.057	0.001603495
58	36,120	0.058	0.001619092
59	36,460	0.059	0.001634688
60	36,800	0.060	0.001650282

and

Table 7. ABAQUS results in compressive strength.

Compressive Strength	Displacement	Maximum Load	Stress
20	5.61	282.24	2.01155
21	5.79	292.32	2.11429
22	5.86	302.40	2.21578
23	5.95	312.48	2.31714
24	6.12	322.56	2.41853
25	6.34	342.72	2.51992
26	6.55	352.80	2.60209
27	6.76	372.96	2.71764
28	6.88	383.04	2.8199
29	7.09	393.12	2.92143
30	7.33	413.28	3.02286
31	7.55	423.36	3.12806
32	7.74	433.44	3.22887
33	7.94	453.60	3.32969
34	8.13	463.68	3.42968
35	8.22	473.76	3.53135
36	8.32	483.84	3.62869
37	8.53	504.00	3.73805
38	8.69	514.08	3.83383
39	8.89	524.16	3.93109
40	9.02	544.32	4.03547
41	9.16	554.40	4.1363
42	9.35	564.48	4.23713
43	9.55	574.56	4.33795
44	9.72	594.72	4.43878
45	9.80	604.80	4.53845
46	9.89	614.88	4.64042
47	9.99	624.96	4.74124
48	10.09	635.04	4.84205
49	10.19	645.12	4.94285
50	10.29	655.20	5.03554
51	10.39	665.28	5.14568
52	10.48	675.36	5.24521
53	10.58	685.44	5.34599
54	10.68	695.52	5.44675
55	10.78	705.60	5.54749
56	10.88	715.50	5.65422
57	11.02	727.65	5.74868
58	11.13	734.23	5.84888
59	11.23	745.50	5.95644
60	11.33	756.30	6.05788

. Table 6 presents data on Young’s modulus, yield stress, and cracking strain based on experimental values categorized by material properties, according to compressive strength. Table 7 shows the values of displacement, maximum load, and stress according to compressive strength based on the FEA analysis results. A two-stage SVR prediction model was constructed based on this data. For the first stage of model training, a total of 40 experimental and interpretive data were used, with cracking strain, Young’s modulus, and yield stress as input variables and compressive strength as the output variable. The collected data were normalized to the range [0, 1] using min-max scaling, and the entire dataset was divided into training and validation sets in an 80:20 ratio. In addition, to improve model generalization and reduce overfitting due to the limited dataset, K-fold cross-validation (K = 5) was applied during the training and evaluation of the SVR model. This method allowed the model to be validated on multiple data subsets, ensuring more reliable performance assessment.

The first-stage prediction model was set with a linear kernel (kernel = “linear”) and a hyperparameter C = 10. The model predicted the load intensity values based on Young’s modulus, yield stress, and cracking strain according to the compressive strength. Test results showed that the MSE was approximately 3.31.

A new dataset was created along with the principal stress and maximum load based on the predicted intensity values from the first stage. The SVR model in stage 2 was trained with the predicted compressive strength as an additional input variable, maintaining the same linear kernel and C = 10. The final model was trained to predict displacement, a representative response variable of structural behavior. Test results showed that the MSE was approximately 0.0151, significantly improving the prediction performance compared to the first stage.

4.3. Prediction Result

Figure 12 shows the comparison results between the displacement, maximum load, and stress values predicted by the proposed two-stage SVR model and the actual analysis values. Each graph compares the analytical values with the predicted values, and the bar graph at the bottom presents the prediction error rate (Error %). The displacement prediction results (Figure 12a) showed a trend where the analytical values were similar to the actual values across all intervals, with the error rate mostly occurring below 3%. In particular, the error decreased to less than 1% in the range around a compressive strength of 35 MPa, confirming the model’s precise prediction performance. The maximum load prediction results in Figure 12b also show a high correlation with the actual values, demonstrating excellent prediction accuracy. Generally, the error rate was maintained within 2%, and relatively large errors (up to 7.2%) were observed only in the 20–25 MPa range. In the stress prediction results (Figure 12c), the SVR model effectively learned the actual stress values, showing a low error rate of less than 3% in most intervals. The error rate distribution was stable, especially in the 30–50 MPa range, and the highest prediction accuracy was observed in the mid-range compressive strength. These results suggest that the SVR model is particularly effective in the 30–50 MPa range, where structural behavior tends to be more stable and less influenced by nonlinearities or boundary effects. The slightly higher error rates observed at 20–25 MPa ranges may be due to greater variability in material behavior and heightened sensitivity to boundary conditions. This indicates that the model’s learning accuracy improves when the training data exhibit more consistent mechanical behavior, as is often the case in moderate- to high-strength concretes. Additionally, a comparison with FEA results confirmed that the SVR model can predict complex finite element analysis results with high accuracy. As a result, the proposed SVR model can reduce calculation time and resources by providing material properties in the design process of precast concrete structures using the open-type joint method.

5. Conclusions

The structural performance of the transverse joint of a PS using the open-type joint method was analyzed in this study. The objective of this study was to verify the effectiveness of the new connection method by comparing it with the traditional CIP method. The load–displacement relationship, crack occurrence patterns, and the strain of the rebar through a four-point loading test were analyzed through a structural experiment. A finite element analysis was performed using ABAQUS for a three-dimensional nonlinear analysis to verify the experimental results and accurately reproduce the bending behavior. Subsequently, as a machine learning model, the SVR model was proposed based on the four-point bending test and analysis data, predicting the material properties of concrete. The conclusions drawn from this study are as follows.

• PS-27 had a yield load measured at 480 kN and a maximum load of 600 kN, showing the lowest structural performance among the specimens. The yield load and the maximum load of CS-27 were measured at 500 kN and 710 kN, respectively. The yield load of PS-40 was 550 kN, and the maximum load was measured at 730 kN. PS-40 experienced relatively more deflection, but the structural behavior in the maximum load interval was similar to that of CS-27. Therefore, if the strength of the non-shrinkage concrete is 40 MPa or higher, the PS joint using the open-type joint method can achieve a structural performance level similar to that of the slab joint using the CIP method.
• CS-B-27 (test) showed the highest strain of approximately 3000 με when the load was 480 kN. After the rebar yielded, the strain increased sharply, resulting in somewhat unfavorable behavior in terms of structural stability. In contrast, PS-B-40 (test) showed the lowest strain of approximately 1100 με when the load was 480 kN, indicating relatively excellent deformation resistance performance under load. The strain measured at the top of the joint of the PS was found to be a maximum of 25 με or less in all specimens, confirming that the joint behaves structurally as a single entity. Therefore, PS-40, which applies the open-type joint method, can achieve structural performance similar to CS-27 in terms of structural and joint performance.
• As a result of the DIC analysis, the strain of PS-40 was evenly distributed across the slab. In contrast, in CS-27, the strain was locally concentrated in the center and increased sharply. Furthermore, in the final failure mode, PS-40 exhibited relatively fewer cracks compared to CS-27. Therefore, the PS-40, which applies the open-type joint method, has superior load distribution capability compared to the CS-27 and can secure better performance in terms of structural stability.
• The load-relative displacement curve of PS-40 (FEA) and the overall crack pattern of PS-40 (FEA) were similar to the results of the four-point loading test, with a relative error in displacement within 5%. Therefore, FEA is a useful method for verifying structural designs because it can accurately predict the structural behavior of the transverse joint of a PS.
• The SVR-based two-stage prediction model was confirmed to predict displacement, maximum load, and stress values with high accuracy using minimal data. In addition, the error rate was reduced to around 3% in the interval beyond 25 MPa. Therefore, SVR can be used as a method to provide material properties during the structural design phase, thereby saving time and costs.