AI-Assisted 3D Modeling Strategy for Microstructure-Based Functional Surfaces Using ChatGPT and Random Forest

Abstract

With the development of computer programming, product design is being carried out through computer programming such as 3D CAD/CAM when developing products, and the cost of developing new products can be greatly reduced by using 3D modeling tools in which various engineering theories are programmed. With the recent rapid development of artificial intelligence (AI), various AI applications have been actively proposed in various fields. In this study, we report a ChatGPT-based 3D modeling program (CMP) which can easily design microstructures, and the theory of wettability that was programmed into the CMP. Since the theory related to wettability is programmed into the CMP, once the user enters basic information about the desired functional surface, such as superhydrophobic surfaces, at the CMP prompt, 3D structures with the performance of superhydrophobicity are immediately provided. Through microstructure fabrication techniques, the microstructure designed by the CMP was actually fabricated, wettability experiments were conducted, and it was confirmed that a superhydrophobic surface was successfully designed. As such, the AI-based CMP can easily design microstructure-based functional surfaces, and we expect this method not only to decrease the fabrication costs of designing micro/nano structures but also to be useful in various fields.

1. Introduction

Today, most product designs and productions heavily utilize computer programs such as 3D CAD (Computer-Aided Design)/CAM (Computer-Aided Manufacturing) and simulation techniques [1,2,3]. They make it possible to predict and visualize the shape and material characteristics of products in advance using pre-programmed engineering theories. Since the engineering theories are programmed, users can simply design by using programmed theories and 3D shape modeling techniques. From a manufacturing cost perspective it has great advantages as it saves many costs such as the time required for design and the cost required for production. However, in order to utilize the programmed engineering theories, the users must have a deep understanding of the relevant engineering knowledge as well.

Recently, with the rapid development of artificial intelligence (AI), applications utilizing AI have been actively introduced in various research fields [4,5,6,7,8]. For example, if users input a simple question into an AI-based program, they can quickly get a related answer, and it has reached the level of not only grammar checking but also creating new text and pictures. This AI technology is constantly developing and through data learning it can provide higher-level results as the number of users increases. These AI-based programs have the advantages of (1) allowing users to quickly obtain results, and (2) allowing anyone to easily use the applications by simply entering a simple command without any related knowledge.

In this study, we introduce an easy design method for functional surfaces based on microstructures using ChatGPT-4, one of the conversational AIs, and a 3D modeling tool. In this study, various theories related to wettability (Young’s equation, Cassie–Baxter/Wenzel state, Laplace pressure, etc.) were programmed into the 3D modeling tool, and the microstructure information reported in various references related to wettability was extracted and trained through AI [9,10,11,12,13,14,15]. Surface texture and complexity play a critical role in determining contact angles, serving as key design elements for controlling superhydrophobicity and wettability [16]. Random forest was used to learn parameters and optimize the design process [17,18]. Other AI approaches, such as artificial neural networks (ANN), also provide methodologies for predicting wettability [19].

Based on these data, the CMP can quickly design 3D structures related to wettability by simply entering information because ChatGPT was used. To verify the performance of the CMP, the microstructure designed by the CMP with information in terms of superhydrophobic surfaces was actually fabricated, and wettability tests were carried out. Through the wettability tests, it was confirmed that the structure designed by the CMP had a superhydrophobic surface. To design a structure related to wettability, a lot of time is required to learn related knowledge; however, the structure can be easily designed by simply entering commands with the CMP.

In this way, the design strategy with the CMP has all the advantages of AI and CAD/CAM: (1) users can quickly obtain the results, (2) anyone can easily use it by simply entering commands without related knowledge, and (3) it can greatly save costs from the perspective of production and design. In this study, we introduced a CMP technique related to wettability and verified its performance and usability through related experiments. We expect that the 3D modeling technique utilizing ChatGPT can be usefully utilized in various fields in the future.

2. Experimental Sections

2.1. Fabrication of Master Mold

The master mold of the microstructures was prepared using photolithography. Initially, to remove moisture on a silicon wafer, the wafer was heated at 120 °C for approximately 10 min on a hot plate. Subsequently, it was uniformly coated in photoresist (PR) (SU8-3000 series, Microchem, Newton, MA, USA) using a spin coater (NSF-100DP, RHABDOS, Seoul, Republic of Korea), followed by a soft baking process to eliminate the solvent. The microstructures were formed using a photomask positioned in an exposure device (MDA-400M, MIDAS SYSTEM, Yongin, Republic of Korea) for UV exposure. To enhance the stability of the exposed photoresist, a post-exposure bake (PEB) was also conducted. The development and rinse processes were performed. Finally, the master mold of the microstructures was successfully prepared.

2.2. Fabrication of Superhydrophobic Surface with PDMS

Using the master mold described in Section 2.1, a superhydrophobic surface was fabricated. Initially, a reverse replica of the master mold was replicated using PFPE (Perfluoropolyether), which incorporates 3 wt% of a photoinitiator suitable for UV curing. The curing was carried out in a UV chamber with an intensity of about 13 mW∙cm⁻² for 8 min, followed by the demolding process. Subsequently, an adequate quantity of PDMS (polydimethylsiloxane) was poured onto the PFPE replica to establish the superhydrophobic surface. The PDMS was prepared using a Sylgard 184 kit (Dow Corning, Midland, MI, USA), mixing the elastomer base and curing agent at a 10:1 wt% ratio. The PDMS mixture was then poured over the PFPE replica, degassed in a vacuum chamber at −1 bar, and cured in an oven at 70 °C for approximately 2 h. Demolding was then carefully implemented to complete the fabrication of the PDMS superhydrophobic surface.

2.3. Measurement

The microstructures on the superhydrophobic surface were inspected using an optical microscope (LV150L, Nikon, Tokyo, Japan) and scanning electron microscope (SEM) (S-4800, Hitachi, Tokyo, Japan). Before the SEM measurement, a metal layer (Pt, <5 nm) was deposited on the surface of the sample to avoid electron charging. To measure the static contact angles (CA), a CA analyzer (Drop Shape Analysis System DSA100 Kruss, Hamburg, Germany) was used, applying ~10 μL drops of deionized water onto the surfaces. The CA measurement was conducted at least five times per sample, with the average values employed for analysis.

3. Results and Discussion

Figure 1 illustrates the complete overview of the ChatGPT-based 3D modeling program (CMP). The 3D modeling tool was developed using wettability theory and ChatGPT. Various references were integrated into the modeling tool, enabling the CMP to design microstructures related to wettability through simple prompt input. Additionally, to verify the performance of the CMP, the microstructures designed by the CMP were fabricated, and wettability tests were conducted. These procedures will be described in detail in Section 3.

3.1. Wetting Transition Theory

To evaluate the pressure exerted on water above a surface featuring PDMS microstructures, the calculation begins by applying the equation for the critical pressure needed to maintain the Cassie state [20]. The critical pressure (P_critical) is determined using the cross-sectional area A of the pillars and their perimeter L [21].

$$P_{critical}=-{\displaystyle \frac{{\gamma }_{LV}fcos{\theta }_{water}}{\left(1-f\right)\left(\frac{A}{L}\right)}}$$

(1)

where γ_LV represents the surface tension between the liquid and air, θ_water is the contact angle, and f is the area fraction. In this case, the contact angle and the surface energy of water used were 110° and 72 mJ∙m⁻², respectively [22]. In this study, two kinds of micro-pillars were considered, with diameters of 20 μm and 40 μm. Additionally, an aspect ratio of two was used to prevent contact between the droplet and the base of the surface [23]. The volume was assumed to be constant and was calculated as outlined in Figure S1.

Based on Equation (1), Figure 2a shows that a reduction in diameters corresponds to a relative increase in critical pressure. This trend is attributed to the fact that a smaller spacing-to-diameter ratio, s/d, results in the denser packing of particles, which enhances their interactions and consequently raises the critical pressure, particularly due to the narrower gaps between particles with smaller diameters. Furthermore, the internal pressure within a droplet can be calculated based on the droplet’s volume (V) of 10 μL, the density of water (ρ) at 1000 kg∙m⁻³, the gravitational constant (g) at 9.81 m∙s⁻²,the apparent contact angle (θ), and the contact radius (R) [24,25].

$$P_{internal}={\displaystyle \frac{V\rho g}{A_{contact}}}={\displaystyle \frac{V\rho g}{\pi R^2}}$$

(2)

When a liquid is dropped onto a solid surface and remains either stationary or in equilibrium, the significant factors include the liquid’s weight (W = Vρg) and the area of contact (A_contact). Assuming that the area of contact forms a circle with radius R₁ = R₂ = R, the configuration of the water droplet and its interaction with the surface depend on the initial volume and surface area, as well as the contact angle, which influences the radius. A higher contact angle leads to a more pointed droplet shape, reducing the contact area with the surface. In contrast, a lower contact angle results in a flatter droplet, increasing the contact area. Thus, the internal pressure related to the contact radius shows that an increase in the radius corresponds to a decrease in internal pressure.

In Figure 2b, as the ratio s/d increases, the contact radius (R) initially decreases, leading to a rise in the critical contact angle (CCA) and a subsequent increase in the internal pressure (P_i). The contact state between the droplet and the surface is determined by comparing two pressures, P_i and P_c. The stability of the liquid droplet, whether formed or sustained on the surface, is assessed by whether (P_i) exceeds (P_c) for the Wenzel state (P_i > P_c) or remains below P_c for the Cassie state (P_i < P_c) (Figure 3).

3.2. Modified Wettability Theory

Typically, the wetting transition is affected by the period of the microstructures, and water droplets on hydrophobic microstructures exhibit a meniscus with a convex downward shape (Figure S2) [26,27]. A key aspect of superhydrophobic design is determining the wetting transition point, where internal pressure exceeds critical pressure. The samples designed using the CMP, based on the wettability theory introduced in Section 3.1, showed a discrepancy between theoretical and experimental values for a specific period. While the CMP-generated samples were expected to display superhydrophobic properties, the actual experimental results confirmed that a transition occurred (Figure S3).

These findings suggested that the relationship between internal and critical pressures was not accurately explained by the existing theory. The error was presumed to be due to the geometric characteristics of the surface structure or inaccuracies in the internal pressure calculation formula. Therefore, to ensure reliable superhydrophobic surface design, a safety factor was introduced into the wettability theory.

The sagging (δ) of water droplets placed on the microstructure is related to the period, which has a crucial role in the transition from the Cassie–Baxter state to the Wenzel state. This can be expressed as follows [28],

$$\mathsf{\delta }={\displaystyle \frac{{\left(\sqrt{2}period-d\right)}^2}{R}}={\displaystyle \frac{{\beta }^2}{R}}$$

(3)

where β represents the sagging effect based on the period and diameter of the microstructure, and R is the radius of the droplet. This equation indicates that the sagging (δ) of the water droplet increases as the period of the microstructure increases.

Using the above equations, the internal pressure P_i can be expressed as follows,

$${\mathrm{P}}_{\mathrm{i}}={\displaystyle \frac{\rho gV{\delta }^2}{\pi {\beta }^4}}$$

(4)

Since the wetting transition occurs when the internal pressure exceeds the critical pressure, a safety factor (n) was applied to the internal pressure, and the modified internal pressure is used.

$${\left({\mathrm{P}}_{\mathrm{i}}\right)}_{modified}={\displaystyle \frac{{\mathrm{P}}_{\mathrm{i}}}{n}}$$

(5)

Thus, the CMP was programmed to prevent the wetting transition by reducing internal pressure using the safety factor (Figure S4).

3.3. Programming

3.3.1. Data Preprocessing and Classification

The method of designing with the 3D modeling tool (Blender 3.5), based on the design function values specified in the prompt, is depicted in Figure 4.

In the data extraction preparation phase, the data analysis program was used to transform graphical data into numerical data [29]. This process used a selection of experimental values from relevant references, summarized in

Table 1. Samples information used for data learning.

Reference	d [μm]	h [μm]	s [μm]
Park et al. [21]	5	5~25	5~50
Gao et al. [30]	25	45	20~150
	35		20~170
	45		20~180
	55		20~210
Lee et al. [31]	105	150	105~420
Yeo et al. [32]	30	50	18.5
	40		18.7
	50		18.6

[21,30,31,32].

Five specific criteria for experiments were established: The first excludes cases where changes in material properties result from surface treatment methods. The second criterion focuses on pillar-patterning surfaces, considering that changes in geometric structures significantly affect outcomes.

The third criterion relies solely on static contact angle measurement data. Fourthly, the interactions of PDMS surfaces are considered only with water (deionized), as the contact angle increases with surface tension [33]. Lastly, the volume of the droplets measured was confined to within 10 μL.

For Dataframe configuration, using the modified wettability theory, a single target variable representing the design variables was extracted. Subsequently, the random forest (RF) machine learning process was applied using these variables since the RF algorithm has been well known as a versatile classification and regression method [17].

Machine learning tasks were conducted using Python (version 3.7.12). The features were linked to the design variables of the microstructures, and the RF algorithm was applied. This method involves iterating decision trees randomly to generate multiple decision trees. The dataset utilizes a test size of 20% (test_size = 0.2), with reproducibility ensured by consistently splitting the data using a seed value set at random_state = 42. To mitigate negative impacts on model training due to very small or large feature values, the StandardScaler function was used to standardize the range of all feature values [18]. The performance of the RF model has been enhanced through various parameter combinations.

3.3.2. ChatGPT-Based Modeling Program (CMP)

This section describes the development of ChatGPT interactions based on structured prompts, utilizing a strategic approach known as ‘prompt engineering’. For high accuracy, the GPT-4 Korean and GPT-4 English models were utilized, achieving accuracies of 77% and 85.5%, respectively. The strategy employs English queries and responses to take advantage of the higher accuracy of the English model [34]. An output template stating “Let’s think step by step” is strategically positioned at the end of statements due to its 78.7% accuracy rate [35]. The prompt pattern includes a Persona pattern, designed to assign specific roles, thereby making explanations accessible to non-experts or learners [36]. As a result, the input message structure merges the Persona pattern (“You are a researcher fabricating superhydrophobic surfaces”.) with the user’s input message and concludes with “Let’s think step by step”. Figure S5 illustrates the user interface in Blender for interactions with ChatGPT, which automates specific calculations or modeling tasks based on user input. This interface processes user inputs, extracts necessary information, and utilizes it as inputs for other operations.

Prompts that include parameters such as diameter, contact angle, and surface state are configured to facilitate modeling or generate textual results. Functions have been devised to process responses and extract essential information efficiently. Regular expressions are employed to retrieve values from responses that are fixed in format, enabling interactions based on these values. A specific programming of the CMP was described in the Supporting Information in detail.

3.4. Validation

To verify the performance of the CMP, the superhydrophobic surfaces which were designed by the CMP were fabricated.

The superhydrophobic surfaces composed of PDMS were replicated from the master mold in Section 2.2, and the contact angle (CA) measurements were carried out.

Figure 5a graphically illustrates the modeling result by the CMP. The aspect ratio of micro-pillar was fixed as two, and the diameters of 20 μm and 40 μm were utilized. The CMP designed the sample with the appropriate period for superhydrophobicity according to the contact angle, which is an input value.

Figure 5 presents microscope images and contact angle measurements of the surfaces. Figure 5b(i,ii) show the microscopic images of the sample with a diameter of 20 μm. In this case, to make the superhydrophobic surfaces the CMP designed the sample with periods of 53 μm and 67 μm. Through the wettability test, it was confirmed that both samples generated by the CMP exhibited superhydrophobicity, with contact angles of 155.2° and 157.3° (Figure 5b(iii,iv)).

To enhance the validation, an additional test was conducted with a diameter of 40 μm (Figure 5c). Figure 5c(i,ii) show the microscopic images of the sample with a diameter of 40 μm. In this case, to achieve the superhydrophobicity, the CMP generated the sample with periods of 115 μm and 200 μm. As shown in Figure 5c(iii,iv), both samples also exhibited superhydrophobicity, with contact angles of 150.5° and 158.2°.

Thus, through the wettability tests, superhydrophobic surfaces were observed in all the samples designed by the CMP, and the performance of the CMP was successfully verified.

4. Conclusions

This research utilized the ChatGPT-based Modeling Program (CMP), optimized through RF machine learning techniques and 3D modeling tools, to design microstructure-based functional surfaces. The primary focus was on fabricating superhydrophobic surfaces, serving as a test for the validity of the design methodology. By incorporating revised theoretical formulas that align with existing wettability research and real experimental conditions, the study identified essential parameters and conditions for developing stable superhydrophobic surfaces and applied these findings through the CMP. The adapted theory supported mathematical and statistical modeling via machine learning and data analysis. Wettability experiments, using contact angle measurements, verified that microstructure surfaces specifically designed to achieve superhydrophobicity with the CMP maintained a stable state with contact angles exceeding 150 degrees. The microstructures designed through the CMP have been experimentally validated, confirming their practicality. This approach extends beyond microstructure fabrication to diverse applications, including product design, optimization, and mass production processes. Additionally, the CMP simplifies the design process by transforming complex microstructure design challenges into simple user queries, enabling non-experts to easily create designs. This demonstrates a clear distinction from traditional methods that require advanced CAD/CAM software or specialized expertise. Such AI-based modeling technology proposes an innovative workflow, promising significant contributions to design and optimization. Lastly, we believe that this study not only gives insights into AI-assisted manufacturing but also suggests the research expansion of micro/nano engineering.

Future Directions

This study focused on designing superhydrophobic surfaces using cylindrical micro-pillars. Future directions include validating more complex geometries and, separately, proposing deep learning-based measurement methods using smartphones and USB microscopes as a cost-effective and accessible alternative [37,38].