Fabrication of Silicon Monolithic Tubes with High-Aspect-Ratio Cavities Using Langmuir–Blodgett Colloidal Lithography and Bayesian-Optimized Continuous Plasma Etching Authors: Artem Osipov, Alina Fumina, Ilya Belyanov, Anastasiya Speshilova, Ekaterina Endiiarova, Alexandr Gerdt Published: 2026 ABSTRACT The fabrication of semiconductor devices using submicron- and nanometer-scale silicon structures is based on lithography (patterning on a substrate) and etching (transferring a pattern onto the substrate) technologies. These processes typically require complex and expensive equipment, as well as extensive experimental optimization of etching parameters, especially for structures with an aspect ratio greater than 10. This work demonstrates a productive and relatively simple approach for fabricating an ordered array of vertically oriented monolithic silicon tubes with high-aspect-ratio internal cavities. The creation of these structures is based on Langmuir–Blodgett colloidal lithography, vacuum magnetron sputtering, and a continuous plasma etching process performed at “room” temperature of the substrate holder. To develop the fabrication process for high-aspect-ratio (>10) structures, we employed Bayesian optimization (a machine learning method), which proved highly efficient in reducing the number of experiments compared to a full factorial analysis. The resulting silicon tubes exhibited an average total reflectance of 3.7% over the 200–1600 nm wavelength range. Introduction Silicon planar solar cells form the basis of modern photovoltaics, which is attributed to their well-established manufacturing technology and the high stability of finished devices. However, a fundamental limitation for their further development is the indirect bandgap of silicon. This implies the spatial separation of the valence band maximum and the conduction band minimum, making direct optical transitions impossible without the participation of a third particle (a phonon). A key consequence is a low absorption coefficient near the silicon absorption edge (λ > 900 nm), which forces the use of relatively thick wafers (150–200 μm). This inevitably increases the cost of photovoltaic converters and narrows the range of their potential applications. Moreover, increasing the substrate thickness lengthens the diffusion path of minority charge carriers to the contacts, raising the probability of their nonradiative recombination before they reach the p–n junction. Consequently, in a planar design, an attempt to increase the short-circuit current by thickening the substrate leads to a drop in the open-circuit voltage and fill factor, thereby limiting the overall conversion efficiency. According to the original Shockley–Queisser detailed balance model, which is based on the key assumption that exclusively radiative recombination dominates, the theoretical efficiency limit for silicon’s bandgap (∼1.12 eV) is about 33%. However, incorporating the Auger recombination mechanism into the model clarifies this value to 29.4%. A promising approach to overcome these limitations is the transition from a planar architecture to arrays of vertically oriented micro- and nanostructures (pillars) that form numerous radial p–n junctions. − In such systems, efficient light absorption is achieved by a combination of optical effects, including multiple scattering, the formation of an effective refractive index gradient, interaction of incident light with waveguide photon modes, and excitation of optical resonances. The combined action of these mechanisms significantly increases the effective optical path of photons, thereby enhancing the absorption coefficient even in thin semiconductor layers. Furthermore, a key advantage of this geometry is the orthogonalization of the directions of light absorption and charge carrier separation. While radiation is absorbed primarily along the vertical axis of the structures, minority carriers need to diffuse only a short transverse distance (comparable to the radius of structure) to reach the depletion region of the radial p–n junction. This decoupling of processes eliminates the traditional dependence of carrier recombination probability on photon penetration depth. Unlike planar elements, where the carrier diffusion length must exceed the absorber layer thickness for efficient collection, this approach enables the use of materials with a short minority carrier diffusion length. Given the above, developing geometries that optimize the processes of light absorption and charge carrier collection is highly relevant. In this regard, ordered arrays of vertically oriented silicon tubes with submicron dimensions are of particular interest. In such structures, an additional increase in the effective p–n junction area is achieved, as the emitter layer is deposited on both the outer and inner surfaces of the tubes. Certainly, any increase in specific surface area (for both tubes and pillars) compared to a planar substrate leads to more surface defects. They act as nonradiative recombination centers and reduce the minority carrier lifetime. However, an important advantage of tubes over pillars is their lower reflectivity across a broad spectral range, as demonstrated in several experimental studies. − This is due to a smoother gradient of the effective refractive index between air and the substrate. Furthermore, the tubular geometry promotes multiple internal reflections of light within their cavities. Thus, the loss of efficiency from a reduced open-circuit voltage and fill factor can be compensated by a significant rise in the short-circuit current. It is worth noting that silicon tubes have great potential not only in photovoltaics but also in mid-infrared photonics. As demonstrated in one study, such arrays can act as metasurfaces and serve as efficient broadband electromagnetic absorbers in the 3–7 μm range. The low atmospheric absorption in the 3–5 μm window makes these structures particularly promising for photodetection devices, remote sensing, and wireless optical communication systems. To date, researchers have used a combination of various lithography and etching methods to fabricate arrays of silicon vertically oriented structures. Specifically, laser interference lithography, electron-beam lithography, , nanoimprint lithography, , colloidal lithography ,, have been employed, along with plasma etching processes such as Bosch, ,− Cryo, , CORE, − and continuous mode at “room” temperature of the substrate holder. For forming hexagonally ordered structures, colloidal lithography is particularly significant, as it does not require expensive equipment and offers high productivity. The most promising direction in this field is considered to be the self-assembly of colloidal particles at the air–water interface with subsequent monolayer transfer onto a substrate. There are two main approaches to implementing this method: 1) self-organization initiated solely by the Marangoni effect; 2) mechanical compression of the monolayer into a state of two-dimensional close packing using the Langmuir–Blodgett technique. An interesting realization of the compression-free approach was proposed by the authors of work, where a confining ring was used to localize the monolayer formation area. Transfer onto the substrate was achieved by immersing it under the monolayer followed by a gentle draining of water. According to the micrographs presented in work, the maximum defect-free domain area did not exceed 120 μm2 for 300 nm polystyrene spheres and 500 μm2 for 1000 nm polystyrene spheres. For comparison, the application of the Langmuir–Blodgett technique in our previous work enabled to obtain monolayers on 76 mm substrates with average defect-free domain areas of 16000 μm2 for 1.25 μm polystyrene spheres and 23000 μm2 for 1.8 spheres (the maximum area reached 1.2 mm2). Thus, the key disadvantage of the first group of methods is the lack of control over the packing density of the spherical particles. This hinders the formation of large-scale defect-free monolayers, as a sparse structure is more susceptible to disruption during the transfer onto the substrate. In existing studies where the colloidal lithography method was used to create a mask for subsequent tube fabrication, etching was performed using the Bosch process. ,, As is known, its cyclical nature leads to the formation of scalloped sidewalls, which is unacceptable for applications requiring minimal carrier recombination losses. In turn, cryogenic etching requires operation at extremely low temperatures and uses oxygen-containing plasma, which imposes serious limitations on the choice of masking materials and complicates the entire process. Specifically, in this case, polystyrene spheres, most commonly used in colloidal lithography, cannot be used due to their susceptibility to oxygen etching. The essence of the CORE method lies in the cyclical repetition of 4 sequential steps. The first step (E) is carried out in SF6 plasma for silicon etching. The second step (C) is an oxygen purge to remove residual fluorine from the reactor. The third step (O) is the passivation of the trench sidewalls in oxygen plasma. The fourth step (R) is necessary to remove the oxide layer from the bottom of the silicon etching windows in SF6 plasma without damaging the oxide layer on the sidewalls. The final step is conducted at minimum pressure and maximum bias voltage. Due to the duration of the E (up to 80 s) and R (up to 35 s) steps, it is possible to obtain smoother surfaces compared to the Bosch process. The key drawbacks of this approach are low etching rate due to the use of only a capacitive plasma source (low density compared to inductively coupled sources (ICP)) and the use of oxygen as a process inhibitor gas (imposes a limitation on organic masks such as photoresist, polystyrene spheres, and others). Another widely used technique for the fabrication of vertically oriented silicon structures is metal-assisted chemical etching (MACE) combined with colloidal lithography. , The essence of MACE is the anisotropic etching of silicon under a metal catalyst in an aqueous HF/H2O2 solution. On the one hand, this method represents a cost-effective approach that does not require complex technological equipment. However, it is worth noting that the use of colloidal lithography requires an operation to reduce the spheres in oxygen plasma. Moreover, creating a high-quality array of vertically oriented silicon structures involves two steps: applying an adhesion layer and depositing the metal catalyst using vacuum deposition methods. Undoubtedly, with a good metal catalyst coating and properly selected silicon etching parameters in the HF/H2O2 solution, it is possible to produce high-quality vertically oriented silicon pillars with high aspect ratio. However, it is not possible to create vertically oriented silicon tubes without performing an additional lithographic operation, especially when using colloidal lithography. Thus, despite progress in fabricating silicon micro- and nanotube arrays, a key challenge is to develop an etching process that is compatible with colloidal lithography and produces structures with smooth sidewalls and a controlled profile. A promising solution may be the use of continuous plasma chemical etching (PCE) at “room” temperature with SF6/C4F8 gas mixture. As demonstrated in several studies, − the geometric parameters of vertically oriented silicon structures (lateral size, height, sidewall angle, period and degree of ordering) directly influence the key characteristics of photovoltaic and photonic devices. The fabrication of such structures involves three main stages of geometry control. First, at the colloidal lithography step, the sphere diameter defines the lattice period of the array, while the monolayer quality determines the packing density and spatial organization. The tube diameter is controlled by subsequently reducing the colloidal particle size in an oxygen plasma. In this process, the external diameter is specified by the sphere size, and the internal diameter is set by the contact area between the sphere and the substrate. Finally, PCE of silicon in a fluorine-based plasma completes the geometry formation, determining the final profile and aspect ratio of the structures. Traditional optimization of such multistage process is based on sequentially varying individual technological parameters while keeping the others fixed. Although this approach helps to identify the physicochemical regularities, it often fails to achieve a really optimal regime, as a checking all parameter values is impractical due to excessive time and resource costs. A promising alternative in this case is the use of machine learning methods, particularly Bayesian optimization. , This approach enables efficient exploration of a multidimensional parameter space without requiring a large initial database. The algorithm iteratively learns from each experiment, constructing a probabilistic model of the objective function and proposing the most promising parameter sets to test in order to achieve the target geometry. Thus, this strategy significantly accelerates the identification of optimal process conditions. Despite the growing interest in machine learning methods, a literature analysis reveals that Bayesian optimization has not yet been applied to develop a fabrication process for vertically oriented silicon tubes. This highlights the significance of our study, which proposes the following approach: 1) use Bayesian optimization to identify the optimal technological parameters for the continuous PCE of silicon at “room” temperature using a mask formed by colloidal lithography; 2) apply multivariate regression analysis to assess the influence of the most significant parameters on the key characteristics of resulting structures (aspect ratio, undercut, inverse sidewall angle, and etching isotropy); 3) measure total reflection coefficients to establish the relationship between the geometric and optical properties of the fabricated silicon structures. Methods Fabrication Process and Equipment The fabrication of vertically oriented monolithic silicon tubes involved seven main stages (Figure ). At Step 1, a mask pattern was created using Langmuir–Blodgett colloidal lithography (Figure a). This process was conducted using an original automated NSL-LB-1 (NSL – nanosphere lithography, LB – Langmuir–Blodgett) system (designed and created by the authors, Russia) with a single barrier (Figure a). As is known, precise control of the monolayer surface pressure is crucial for fabricating defect-free coatings. Consequently, the NSL-LB-1 was equipped with a feedback control system to maintain a constant specified pressure value throughout the entire process. The monolayer surface pressure at any given time was calculated as the difference between the surface tension of pure deionized water (close to 72.8 mN/m) and the current surface tension of the subphase with the monolayer. These readings were obtained with a sensitivity of 0.05 mN/m using a sensor that employed a laser beam to track the deflection of a measuring plate connected to a Wilhelmy plate. Scheme of step-by-step manufacturing of monolithic vertically oriented silicon tubes: (a) colloidal lithography process; (b) reduction of sphere size in an oxygen plasma; (c) formation of silicon pillars using PCE process; (d) magnetron deposition of chromium; (e) removing latex spheres; (f) formation of cavities using PCE process; (g) removing chromium using wet etching. (a) General view of the Langmuir–Blodgett colloidal lithography NSL-LB-1 system; (b) general view of the ICP Plasma Etcher 100 system; (c) general view of the magnetron sputtering DC MS-76 system. Steps 2 and 3 were performed using an original ICP Plasma Etcher 100 system (designed and created by the authors, Russia) with a cylindrical inductively coupled plasma source (Figure b). First, the sphere size was reduced in an oxygen plasma to define the outer diameter of silicon structures (Figure b). This was followed by the continuous etching of vertically oriented pillars in SF6/C4F8 plasma at “room” temperature of the substrate holder (Figure c). At Step 4, the fabricated structures were coated with a thin chromium layer (Figure d) using a custom-built DC MS-76 (DC – direct current, MS – magnetron sputtering) system (designed and created by the authors, Russia) (Figure c). Step 5 involved the removal of the spheres in deionized water (Figure e) using an ultrasonic bath (model GS0915, Granbo, China). As a result, etching windows were formed on the top surfaces of the silicon pillars without the need for expensive lithographic equipment with an alignment function. During Step 6, the continuous plasma chemical etching in SF6/C4F8 plasma at “room” temperature of the substrate holder was repeated. It was used to form cavities within the pillars, thereby creating the tubular structures (Figure f). Finally, at the Step 7, the chromium mask was selectively removed in a wet etchant (Figure g). Materials and Reagents In this work, commercial polystyrene microspheres with a diameter of 1.8 μm were used. They were supplied as an aqueous suspension with a solid content of 10% from the Institute of High Molecular Compounds of the Russian Academy of Sciences (Russia). The uniformity in the size of the microspheres was confirmed using a laser particle analyzer (model Laska-TD, Biomedical Systems LLC, Russia), which showed a coefficient of variation of less than 2.5%. The substrates used were single-sided polished monocrystalline silicon wafers (n-type, resistivity of 0.004 Ω·cm) with a (100) orientation, diameter of 76 mm and thickness of 380 μm (purchased from Silicon LLC, Russia). Prior to use, the substrates were cleaned by boiling in acetone, after which they were treated in a piranha solution (1:3 H2O2:H2SO4) for 45 min at 120 °C, followed by washing in deionized water three times. All chemicals used (acetone, sulfuric acid, hydrogen peroxide, ethyl alcohol, sodium hydroxide) were of analytical grade (purchased from SigmaTek LLC, Russia). The deionized water had a resistivity of 10.6 MΩ·cm at 24 °C, corresponding to an “ultrapure” grade. The latex spheres were etched in an oxygen plasma (O2, class 6.0, 99.9999%, GOST TU 2114-001-05798345-2007) to reduce their size, thereby defining a mask pattern with the required submicron dimensions for subsequent silicon structuring. Silicon was then etched in SF6/C4F8 gas mixture: sulfur hexafluoride (SF6, 99.998%, GOST TU 6-02-1249-83) was used as the primary etching gas, and octafluorocyclobutane (C4F8, GOST TU 2412-128-05807960-96) was employed as the passivating agent. Data Analysis Following the process of cavity etching, the resulting structures were analyzed. The etching depth (H2), the mask undercut (B), and the sidewall angles of the etching profile (∠α and ∠β) were evaluated from micrographs obtained using a Carl Zeiss (Supra 55VP, Germany) scanning electron microscope (SEM) with an accuracy of ±2.5% (Figure ). Scheme for measuring the geometric parameters of Si cavities fabricated using continuous plasma chemical etching in SF6/C4F8 gas mixture: (a) cross-sectional view; (b) top view. Measurement of Optical Characteristics of Plasma and Silicon Structures To investigate the spectral characteristics of plasma, optical emission spectra from the radiation in the reaction chamber were recorded. For this purpose, a spectrometer (model HR 4000, Ocean Optics, USA) with a slit width of 5 μm was used. The optical signal was recorded in the wavelength range of 200–1120 nm. The spectrometer was connected to the plasma etching system using an optical fiber assembly that transmitted the plasma emission to the spectrometer’s entrance slit. The reflectance of the fabricated silicon structures was measured using a spectrophotometer (model UV-3600, Shimadzu, Japan). The instrument was equipped with a combined light source system (a 50 W halogen and deuterium lamps) and three detectors: a photomultiplier tube for the ultraviolet (UV) and visible spectral ranges, as well as InGaAs and PbS semiconductor detectors for the near-infrared (IR) range. All measurements were performed in the wavelength range from 200 to 1600 nm with a scanning speed of 250 nm/min. The monochromator’s spectral slit width was set to 5 nm, with a wavelength resolution of 1 nm. The total light reflected from a sample is the sum of two components: specular and diffuse. Specular reflection, observed at an angle equal to the angle of incidence, occurs on optically smooth surfaces where the surface roughness is significantly smaller than the wavelength of the incident light. It was measured using a standard absolute specular reflectance accessory with incidence angle of 5°. Prior to measurements, the system was calibrated using a reference mirror with a known reflectance close to 100% across the operational spectral range. Diffuse reflection results from the penetration of incident light into the sample volume, where part of the radiation is absorbed, and part undergoes multiple scattering at internal interfaces. In the case of the fabricated silicon structures, the contribution of the diffuse component is primarily due to the disruption of the coherent addition of waves reflected from the surface relief. This occurs when the size and period of the structures become comparable to or exceed the wavelength of the incident light. Diffuse reflectance was measured using an integrating sphere accessory (ISR-3100). The sphere featured a highly reflective (over 99%) internal coating (Spectralon), which ensured multiple rereflections and averaging of the collected light. In this configuration, to prevent the detection of the specular component, the sample was illuminated at a 0° angle (normal incidence). This allowed the specular component to exit the sphere through the entrance port without reaching the detector. A PTFE (polytetrafluoroethylene) standard with a reflectance of 95–99% across the spectral range of interest was used as a reference sample for baseline correction. The obtained experimental spectra were processed and analyzed using the Spectragryph 1.2.14 software. Bayesian Optimization Framework Bayesian optimization was employed to develop a fabrication technology for directional silicon cavities with high aspect ratios (>10) in the submicron size range. The applied methodology is based on Bayes’ theorem, which provides a probabilistic framework for sequentially updating prior assumptions about the objective function as new experimental data are acquired. Bayes’ theorem is formulated as follows: 1 P(M|E)∝P(E|M)×P(M) where P­(M) is the prior probability of hypothesis M (the initial assumption), P­(E|M) is the likelihood (the probability of observing data E given that hypothesis M is true), and P­(M|E) is the posterior probability (the updated assessment of hypothesis M after accounting for data E). In the context of etching process development, the principle of the method involves the sequential updating of a probabilistic model for the response function, which relates the target process characteristics to the varied technological parameters. This study was focused on multiobjective optimization to simultaneously achieve several distinct goals: 1) minimization of the inverse sidewall angle (C, eq ); 2) minimization of the mask undercut (B, eq ); 3) minimization of the etching isotropy (I, eq ); 4) maximization of the structure aspect ratio (AR, eq ). The input variable parameters of the etching process included ICP power (WICP), pressure (P), bias voltage (Ubias), temperature (T), SF6 flow rate (QSF6), and C4F8 flow rate (QC4F8). 2 C=|90−α+β2| 3 B=W2−W12 4 I=BH2 5 AR=H2W1 Since standard Bayesian optimization is applied to functions with a scalar output, several targets had to be combined into a single comprehensive criterion. In our case, this task can be mathematically formulated as finding an input parameter vector x = (x 1, x 2, ..., x 6) that minimizes the objective vector-function F­(x), which is a superposition of four etching output parameters normalized by their mean values: 6 x*=arg⁡minx∈X⁡F(x) 7 F(x)=(f1(x),f2(x),f3(x),f4(x)) where X is the feasible decision space. In general, the goal of multiobjective optimization is to find a set of optimal solutions that represent trade-offs between the different objectives. These are known as Pareto-optimal solutions, where improving one target metric is impossible without degrading at least one other. The Bayesian optimization algorithm is implemented as an iterative process. In the first stage, a prior probabilistic model of the response function is constructed, reflecting the initial assumptions about the dependence of the objective function on the input parameters based on the available experimental data. The initial prior model was built using data from two starting experiments, the parameters of which are presented in Table . 1Values of the Input Parameters in the Initial ExperimentsNo.WICP, W|Ubias|, VP, PaT, °CQSF6, sccmQC4F8, sccm110001252.315103521000502.311332 Values of the Input Parameters in the Initial Experiments This work employed a Gaussian process as the prior model due to its capability for flexible, nonparametric function approximation without requiring a predefined analytical form. Furthermore, it is based on a kernel (covariance function) that assumes the response function’s continuity, implying that similar input parameters provide similar response values. This property aligns well with the behavior of most physicochemical processes, where a smooth variation in technological parameters leads to a predictable and continuous change in output characteristics. The Bayesian optimization was implemented using the GPyOpt library, an open-source Python package. Its core functionality relies on the GPy library for Gaussian process modeling. At each subsequent algorithm iteration:Based on the current model, the acquisition function identifies the most informative point in the parameter space to conduct the next experiment.After performing the experiment with the specified parameters and measuring the response function value, the data set is updated. The model is then refined according to Bayes’ theorem: the current distribution (which serves as the prior for this iteration) is updated with the new data and transformed into the posterior distribution. Based on the current model, the acquisition function identifies the most informative point in the parameter space to conduct the next experiment. Based on the current model, the acquisition function identifies the most informative point in the parameter space to conduct the next experiment. After performing the experiment with the specified parameters and measuring the response function value, the data set is updated. The model is then refined according to Bayes’ theorem: the current distribution (which serves as the prior for this iteration) is updated with the new data and transformed into the posterior distribution. After performing the experiment with the specified parameters and measuring the response function value, the data set is updated. The model is then refined according to Bayes’ theorem: the current distribution (which serves as the prior for this iteration) is updated with the new data and transformed into the posterior distribution. This study employed the Lower Confidence Bound (LCB) as the acquisition function. This criterion is based on the assumption that the best estimate of the objective function’s minimum lies near the lower bound of its confidence interval (CI). It defines the rule for selecting the next experimental point as the one where this estimate is minimized. The LCB calculation relies on a probabilistic model that provides two key metrics for any point in the parameter space: 1) the expected value μ­(x), representing the most probable value of the objective function at that point; 2) the standard deviation σ­(x), representing a statistical measure of the model’s prediction uncertainty. Based on these quantities, a confidence interval of the form [μ­(x) – κσ­(x), μ­(x) + κσ­(x)] is constructed, where κ is a weight parameter that balances the exploration strategy (investigating regions with high uncertainty σ­(x)) and the exploitation strategy (utilizing regions with known low values of μ­(x)). The LCB acquisition function is calculated as 8 LCB(x)=μ(x)−κσ(x) Thus, minimizing the LCB­(x) function enables a targeted exploration of regions where the model predicts either low response values or high uncertainty, which leads to the search for a global minimum. Hyperparameters consist of the probabilistic model’s parameters (length scale and variance) and the acquisition function’s parameter (weight). Their selection significantly influences the algorithm’s behavior, determining properties such as the smoothness of the approximating function, its sensitivity to parameter changes, and the balance between exploring the parameter space and exploiting current knowledge. In this work, the following hyperparameter values were chosen to provide neutral starting conditions for optimization in a limited initial data set: Length scale = 1. This value assumes that significant correlation between data points is maintained on a scale comparable to the range of the normalized input parameters. This corresponds to a prior assumption of a moderate smoothness for the response function.Variance = 1. This value sets a neutral prior estimate for the scale of the response function’s variability, neither underestimating nor overestimating it relative to the scale of the normalized data.Weight = 1. This value establishes a balanced mode for the acquisition function at the initial stage, giving equal priority to the exploration and exploitation strategies. Length scale = 1. This value assumes that significant correlation between data points is maintained on a scale comparable to the range of the normalized input parameters. This corresponds to a prior assumption of a moderate smoothness for the response function. Length scale = 1. This value assumes that significant correlation between data points is maintained on a scale comparable to the range of the normalized input parameters. This corresponds to a prior assumption of a moderate smoothness for the response function. Variance = 1. This value sets a neutral prior estimate for the scale of the response function’s variability, neither underestimating nor overestimating it relative to the scale of the normalized data. Variance = 1. This value sets a neutral prior estimate for the scale of the response function’s variability, neither underestimating nor overestimating it relative to the scale of the normalized data. Weight = 1. This value establishes a balanced mode for the acquisition function at the initial stage, giving equal priority to the exploration and exploitation strategies. Weight = 1. This value establishes a balanced mode for the acquisition function at the initial stage, giving equal priority to the exploration and exploitation strategies. Results and Discussion Bayesian Optimization for Developing Etching Technology of Silicon Cavities To form vertically oriented silicon tubes, it is necessary to develop a technology for etching of high-aspect-ratio nanocavities (AR > 10) in silicon pillars. In this regard, test structures with an array of nanoscale etching windows were fabricated using Langmuir–Blodgett colloidal lithography and vacuum magnetron sputtering of chromium (Figure ). SEM images of the mask fabrication process: (a) deposition of polystyrene spheres using Langmuir–Blodgett colloidal lithography (technological parameters include SP – surface pressure (mN/m), n – concentration of ethyl alcohol in the colloidal suspension (%), V – volume of the deposited portion (ul), v1 – initial monolayer compression rate (mm/min), v2 – monolayer transfer rate to the substrate (mm/min)); (b) reduction of spheres in oxygen plasma; (c) deposition of chromium by vacuum magnetron sputtering and removal of spheres in an ultrasonic bath. To develop the technology of high-aspect-ratio cavity etching, Bayesian optimization was used to minimize time and financial costs by reducing the number of required experiments. The range and step size for the variation of the input parameters are presented in Table . 2Input Parameters of Bayesian OptimizationParameterRange of valuesStep sizeWICP, W500–150075|Ubias|, V25–15015P, Pa2–4.50.15T, °C1–352.5QSF6, sccm7–251QC4F8, sccm20–551 Input Parameters of Bayesian Optimization Table summarizes the input technological parameters and the corresponding output data obtained at each iteration of the Bayesian optimization process. The primary objective was to develop a fabrication process for vertically oriented silicon cavities with an aspect ratio greater than 15, an etching isotropy below 0.05, and a mask undercut not exceeding 350 nm. 3Results of the Multi-Objective Bayesian Optimization of Plasma Chemical Etching Parameters for Fabricating High-Aspect-Ratio Silicon Cavities with Nanoscale Dimensions (Experiment No. 14 Indicates the Optimal Mode, Experiments Nos. 1 and 2 Reflect the Starting Points) Input technological parametersOutput parametersNo.WICP, W|Ubias|, VP, PaT, °CQSF6, sccmQC4F8, sccmC, °B, nmIAR110001252,31510351.45256.70.0396.8621000502,3113322.81223.30.04311.4131000752.3513323.40410.50.02614.914750502.3513323.2756.90.0157.8951250502.3513321.45493.70.03915.93610001252.7112331.82268.80.03513.51710001251.9112330.25217.50.0256.498102540318.510339.00348.60.1424.029950150333.57223.68498.70.05811.5910110040333.58257.18613.90.1237.371110001252.33111340.321120.0288.281211001303.58.514445.352090.0398.181311001502.1513.59242.588750.0637.961410251303.353.524502.3318.40.03615.66 Results of the Multi-Objective Bayesian Optimization of Plasma Chemical Etching Parameters for Fabricating High-Aspect-Ratio Silicon Cavities with Nanoscale Dimensions (Experiment No. 14 Indicates the Optimal Mode, Experiments Nos. 1 and 2 Reflect the Starting Points) The optimization was stopped after 12th iteration (excluding the two initial points), even though the algorithm continued to propose new points for investigation. The reason was that the experimental goals had been achieved: the found regime (Experiment No. 14 in table) fully met the predefined target values for the output characteristics. As a result, an optimal etching mode was identified, which enabled the fabrication of structures with the following parameters: AR ≥ 15, I ≈ 0.036, and B < 320 nm (Figure a,b). In addition, the selectivity of silicon etching with respect to chromium was approximately 490 with an initial mask thickness of ∼60 nm and a mask thickness after etching of ∼42 nm (Figure c). It is worth noting that this solution represents a local optimum (minimum) of objective function within the performed experimental cycle. Therefore, the term “optimal etching mode” in this context should be understood as a practically achievable solution that meets the established technological requirements. SEM images of formed structures in mode No. 14 of Bayesian optimization: (a) top view; (b) cross-sectional view; (c) mask thickness after etching. To verify the correct adaptation of the model during training, an analysis of the dynamics of its key hyperparameters (length scale and signal variance) was conducted. The length scale determines the smoothness of the Gaussian process model by setting the characteristic distance for correlation decay between points in the parameter space. The signal variance defines the expected spread of the target function’s values around its mean. As can be seen from Figure , low values of both hyperparameters were observed at the beginning of the optimization (iterations 1–6 of the algorithm). This indicates that the model was focused on local features of the data, avoiding distant extrapolations. This behavior is consistent with the data from Table , which show that the output characteristics changed significantly even with small variations in the technological in the technological parameters. Therefore, the initial phase served as local exploration to establish a more robust probability distribution around the starting points. Plots of the dependences of hyperparameters on the iteration number. A key shift in the model’s strategy occurred at the seventh iteration, as evidenced by a simultaneous increase in both the length scale and the signal variance. This adaptation was triggered by a sharp deviation of the output parameters from the target optimization directions in the previous step: an increase in the inverse sidewall angle, isotropy, undercut and a decrease in the aspect ratio. This outcome most likely invalidated the model’s initial assumption that the response function exhibited high local variability. Consequently, the model increased the length scale at the seventh iteration. This adjustment reflects a revised hypothesis of a smoother objective function, where the observed poor result is interpreted not as a local outlier but as part of a broader suboptimal region. The concurrent rise in the signal variance indicated an expansion of the expected range of response values, which corresponds to a transition toward exploring more distant areas of the parameter space. A similar model adaptation mechanism was observed at the ninth iteration in response to undesirable extrema discovered after the eighth iteration. Following the strategic shift from local to global exploration, the model successfully located a region with acceptable output characteristic values. Subsequently, it reduced the hyperparameter values and entered a parameter refinement phase aimed at precisely determining the optimum (iterations 10–12). Thus, the evolution of the Gaussian process hyperparameters illustrates how the model balances local and global search. Multivariate Regression Analysis The simultaneous variation of technological parameters during Bayesian optimization results in the mixing of their effects, making traditional pairwise analysis unsuitable for assessing the individual contribution of each factor. To reliably identify potential relationships, it is necessary to use statistical methods, such as multivariate regression, which can estimate the isolated effect of each independent parameter (predictor) on the dependent variable (response) while holding all other parameters constant. In this work, the Ordinary Least Squares (OLS) method, implemented in the Python statsmodels library, was used to construct and analyze the regression models. The algorithm for identifying significant linear and nonlinear relationships for each input parameter is presented in Figure . Algorithm for checking linear and nonlinear effects in a regression model. Based on 14 experimental data points obtained through Bayesian optimization, four independent regression models were constructed. Each model described the relationship between one of the output parameters (inverse sidewall angle, undercut, isotropy, aspect ratio) and the six input factors: X1 – ICP power, X2 – bias voltage, X3 – pressure, X4 – temperature, X5 – SF6 flow rate, X6 – C4F8 flow rate. To avoid overfitting, which is inevitable when using complex models on small data sets, the initial analysis was limited to a linear regression model. In the equation shown in Figure , the intercept (β0) corresponds to the predicted value of the response Y in a hypothetical situation where all model predictors are zero. Given that such a point often lies outside the region of physical feasibility (as in our process), this coefficient lacks a meaningful interpretation. Its primary function is to ensure the adequacy of the approximation by shifting the regression hyperplane to minimize prediction errors. Meaningful insight is provided by the partial regression coefficients (βi) which characterize the expected change in the output parameter Y for a one-unit increase in the corresponding parameter Xi, provided all other variables remain constant. The random error (ε) represents the difference between the value of Y predicted by the model and the actual experimental value. Following the construction of the linear model, a key step is the analysis of partial residual plots for each predictor. These plots visualize the true form of the relationship between Xi and Y after “cleaning” it from the influence of other factors. The x-axis represents the partial residuals of the predictor: the part of Xi that is independent of the other factors in the model. The y-axis represents the partial residuals of the dependent variable Y: the part of Y that could not be explained by the other predictors. To identify hidden trends, a nonparametric LOESS (Locally Estimated Scatterplot Smoothing) line is added to these plots. This line approximates the residual point cloud without assuming a specific mathematical form for the relationship. If the LOESS line is close to a straight line, it suggests a potential linear dependence. To confirm its statistical significance, two conditions must be fulfilled simultaneously: Statistical significance of the coefficient – the p-value for βi must be below a chosen significance level (in this work, a value of 0.1 was selected, corresponding to a 90% confidence interval).High model explanatory power – the adjusted coefficient of determination (this parameter indicates the proportion of the variance in the output variable explained by the model) must exceed 0.7, which is considered a satisfactory level of predictive power in engineering practice. Statistical significance of the coefficient – the p-value for βi must be below a chosen significance level (in this work, a value of 0.1 was selected, corresponding to a 90% confidence interval). Statistical significance of the coefficient – the p-value for βi must be below a chosen significance level (in this work, a value of 0.1 was selected, corresponding to a 90% confidence interval). High model explanatory power – the adjusted coefficient of determination (this parameter indicates the proportion of the variance in the output variable explained by the model) must exceed 0.7, which is considered a satisfactory level of predictive power in engineering practice. High model explanatory power – the adjusted coefficient of determination (this parameter indicates the proportion of the variance in the output variable explained by the model) must exceed 0.7, which is considered a satisfactory level of predictive power in engineering practice. If a nonlinear trend is identified on the LOESS plot, a hypothesis about the presence of a quadratic relationship is put forward. To test it, the original linear model is modified to a second-degree polynomial regression. If several potentially nonlinear effects are detected, the quadratic terms (Xi 2) are added to the model sequentially. A quadratic relationship is considered statistically significant and justified only if the following two conditions are fulfilled simultaneously:Statistical significance of the coefficient – the p-value for βii (the coefficient of the quadratic term) must be less than 0.1.Improvement in model quality – a decrease in the Akaike (AIC) and Bayesian (BIC) information criteria compared to the original linear model must be observed. These parameters evaluate the trade-off between the model’s goodness-of-fit and its complexity: a decrease in their values indicates that the improvement in accuracy achieved by adding the quadratic term statistically justifies the increase in model complexity, confirming its better relative adequacy. Statistical significance of the coefficient – the p-value for βii (the coefficient of the quadratic term) must be less than 0.1. Statistical significance of the coefficient – the p-value for βii (the coefficient of the quadratic term) must be less than 0.1. Improvement in model quality – a decrease in the Akaike (AIC) and Bayesian (BIC) information criteria compared to the original linear model must be observed. These parameters evaluate the trade-off between the model’s goodness-of-fit and its complexity: a decrease in their values indicates that the improvement in accuracy achieved by adding the quadratic term statistically justifies the increase in model complexity, confirming its better relative adequacy. Improvement in model quality – a decrease in the Akaike (AIC) and Bayesian (BIC) information criteria compared to the original linear model must be observed. These parameters evaluate the trade-off between the model’s goodness-of-fit and its complexity: a decrease in their values indicates that the improvement in accuracy achieved by adding the quadratic term statistically justifies the increase in model complexity, confirming its better relative adequacy. As a result of applying the described algorithm to the experimental data, statistically significant quadratic dependencies were identified for the relationship between etching isotropy and SF6 flow rate, and the relationship between aspect ratio and bias voltage. As shown in Figures a­(I) and a­(I), the partial residual plots for the linear models exhibit pronounced nonlinearity, confirmed by the shape of the LOESS (red dashed line). After incorporating the quadratic term into the model, the form of the relationship on the corresponding residual plots becomes nearly linear (Figures a­(II), a­(II)). This change confirms the adequacy of the performed approximation: since the quadratic component responsible for the nonlinearity is now explicitly accounted for in the model, its manifestation in the residuals is minimized. (a) Partial residual plots of the regression model for the dependence of isotropy on SF6 flow rate (I – linear relationship; II – quadratic relationship); averaged prediction plots for the dependence of isotropy on bias voltage magnitude (b) and SF6 flow rate (c). (a) Partial residual plots of the regression model for the dependence of aspect ratio on bias voltage (I – linear relationship; II – quadratic relationship); averaged prediction plots for the dependence of aspect ratio on bias voltage magnitude (b), ICP power (c), SF6 flow rate (d), C4F8 flow rate (e). At the final stage, plots were constructed to visualize the identified significant relationships. These plots display the model-predicted response value Y against the target predictor, while holding other independent variables constant at their mean levels. As can be seen in Figures b,c, b–e, , the experimental data points from the Bayesian optimization are concentrated in regions of extremum search and do not uniformly cover the entire parameter space. Therefore, their role is not to validate the model’s adequacy, but rather to illustrate the original space within which the model operated. The primary purpose of the averaged prediction plot is to clearly represent the isolated effect of a factor on the response, as described by the regression model. The bias voltage had a significant influence on all investigated output characteristics. An increase in its magnitude in the range of 40 to 150 V led to a monotonic decrease in the etching isotropy (Figure b), undercut (Figure a), and inverse sidewall angle (Figure b). These trends are attributed to the increase in ion energy and the narrowing of ion angular distribution function with raising bias voltage magnitude. The nonlinear dependence of the aspect ratio on the bias voltage (Figure b), characterized by a maximum at intermediate parameter values, is most likely explained by the competing effects of two factors. On the one hand, an increase in the bias voltage magnitude raises the proportion of ions moving perpendicular to the substrate, thereby suppressing lateral etching. On the other hand, at high bias voltage magnitudes, ion bombardment intensifies not only on the silicon substrate but also on the surface of the masking layer (chromium). Due to the isotropic nature of magnetron sputtering, the complex-geometry surface (an array of silicon pillars with polystyrene spheres on their tops) was coated by the chromium from all directions. Consequently, at the base of each sphere (in the area of the future etching window edge), a mask with sloped walls and reduced thickness was formed. As a result, intense ion bombardment caused faster damage to these regions, leading to an expansion of the etching window and a decrease in the aspect ratio of structures. Both the experimental data and the regression model indicate that an initial increase in the SF6 content in the gas mixture does not enhance isotropic etching, as one might expect (Figure c). This suggests that at low SF6 flow rates, the predominant mechanism is not the chemical etching of silicon, but the passivation of its surface by a fluorocarbon film. Consequently, the growth of atomic fluorine concentration, resulting from a greater number of SF6 molecules entering the reaction volume, primarily contributes to an increase in the vertical component of the etching rate. This hypothesis is further supported by other observed changes: 1) a decrease in the inverse sidewall angle (Figure f), indicating the formation of a more vertical profile; 2) an increase in the aspect ratio (Figure d), evidencing directional etching without significant lateral expansion of the etching window. In contrast, an inverse relationship is observed for the dependence of the aspect ratio on the C4F8 flow rate (Figure e). Most likely, this is due to the fact that at a constant total gas flow rate, an increase in the C4F8 content elevates the concentration of various fluorocarbon species in plasma, while the concentration of fluorine atoms decreases. This is also supported by optical actinometry and spectroscopy results (Figure ). (a) Typical SF6/C4F8 plasma spectrum; (b) emission spectra of the 251.9 nm line (CF2) at different C4F8 percentages in the total gas mixture; (c) emission spectra of the 516.2 nm line (C2) at different C4F8 percentages in the total gas mixture; (d) concentration of atomic fluorine as a function of the C4F8 percentage in the total gas mixture; (e) emission intensity of the 251.9 nm (CF2) and 516.2 nm (C2) lines as a function of the C4F8 percentage in the total gas mixture. Measurements were performed in the wavelength range of 200 to 1120 nm. To correlate etching parameters with plasma chemical kinetics, in terms of the formation of a fluorocarbon protective film, a 251.9 nm emission line was selected. It corresponds to CF2 and directly influences the intensity of protective fluorocarbon layer formation on the silicon substrates during plasma chemical etching. The most intense carbon emission line (C2) at 516.2 nm was also chosen. The experiments were conducted using the technological parameters previously identified during Bayesian optimization (experiment No. 14 in Table ). Only the C4F8 percentage in the total gas mixture (C4F8/(SF6+C4F8)) was varied from 60% to 80%, while the total gas flow rate was kept constant. As shown in Figure e, an increase in the C4F8 percentage in the total gas mixture from 60% to 80% (while keeping other technological parameters constant), leads to an increase in the emission intensities of CF2 (Figure b) and C2 (Figure c) lines by ∼1.75 and ∼5 times, respectively. Concurrently, the concentration of atomic fluorine decreases by ∼4 times (Figure d). As a result, the rate of passivating film formation on all surfaces, including the bottom, begins to exceed the rate of its removal by ion bombardment. Therefore, the access of atomic fluorine to the silicon surface can be partially blocked, and vertical etching can slow down, which is reflected in the decrease of the structure’s aspect ratio. The generation rate of chemically active particles in plasma is determined by the dissociation frequency of gas molecules upon collisions with electrons. Since the electron concentration is proportional to the applied ICP power, an increase in power can be expected to lead to a higher concentration of atomic fluorine and, consequently, an increase in the silicon chemical etching rate. However, if the gas mixture contains not only the main etchant gas SF6 (the primary source of atomic fluorine) but also gases whose dissociation products can redeposit on the etching surface as a protective polymer film, the etching behavior can change differently in the vertical and lateral directions. This is explained by the competition between two processes: 1) the enhancement of chemical etching due to the increased concentration of atomic fluorine; 2) the enhancement of surface passivation due to the increased concentration of carbon-containing radicals. In our process, the increase in the aspect ratio of the structures with increasing ICP power (Figure c) indicates that the vertical component of the silicon etching rate increases more intensively. At the same time, the increase in mask undercut (Figure e) proves that the lateral etching component is also present and intensifies. Averaged prediction plots for the dependence of undercut on bias voltage magnitude (a), pressure (c), ICP power (e); averaged prediction plots for the dependence of inverse sidewall angle on bias voltage magnitude (b), pressure (d), SF6 flow rate (f). Another significant parameter directly affecting the intensity of gas molecule dissociation is the pressure in the reaction chamber. At low pressure, the large mean free path results in high average kinetic energy and temperature of electrons, which enhances dissociation. At high pressure, conversely, dissociation is suppressed, and the radical concentration decreases. However, raising pressure also increases the residence time of particles in plasma, which can enhance the dissociation efficiency of both SF6 and C4F8 molecules within a certain power range. According to Figure c,d, an increase in pressure from 1.9 to 3.35 Pa was accompanied by a monotonic increase in mask undercut and inverse sidewall angle. This indicates an increase in the lateral component of the etching rate and a deviation of profile from the vertical. It can be assumed that the dominant mechanism in this case was the chemical etching of silicon by fluorine radicals. Their concentration likely increased more significantly compared to carbon-containing radicals due to the increased particle residence time in plasma. Furthermore, an increase in pressure at a fixed value of the bias voltage leads to a decrease in the mean free path of plasma particles, including ions. This leads to a broadening of the ion angular distribution function, contributing to the increase in both the mask undercut and the inverse sidewall angle. It should be clarified that the goal of the regression analysis in this work was not to predict the influence of technological parameters on the output characteristics in an unexplored range of values, but rather to identify patterns within the obtained set of experimental data. Therefore, the focus was placed not on the absolute values of the regression coefficients, but on determining the significance and the nature of the effect of each parameter within the studied range of values. Formation of Silicon Tubes and Investigation of Their Optical Characteristics Previously, we developed a technology for fabricating vertically oriented silicon pillars using polystyrene sphere mask, which was also optimized by Bayesian method (technological parameters: WICP = 1500 W, Ubias = −65 V, P = 3.25 Pa, T = 10 °C, QSF6 = 19 sccm, QC4F8 = 55 sccm). In the present work, we have developed a technology for creating high-aspect-ratio silicon cavities with nanoscale dimensions. Consequently, the next step is the creation of vertically oriented monolithic silicon tubes. The technological process shown in Figure was used to form the tubes. The optimal mode of cavity etching, identified during the Bayesian optimization, was used to create silicon tubes based on pillars of different heights (H1): ∼5.5 μm, ∼4.0 μm, ∼2.5 μm, ∼1.6 μm (Figure a,e,i,m). The center-to-center distance between structures in the arrays was ∼1.8 μm. The average pillar diameter (Wpillar), calculated as the arithmetic mean of the diameters at the base and the top, ranged from ∼1.35 to ∼1.55 μm. Following the magnetron sputtering of chromium through the polystyrene sphere mask and subsequent sphere removal, etching windows with sizes ranging from ∼200 to ∼350 nm were formed (Figure b,f,j,n). The reduction in window size with increasing initial pillar height can be explained by sphere erosion. In other words, during the longer time required to etching of high structures, the spheres undergo more significant size reduction, which decreases their contact area with the silicon. The etching process resulted in cavities with the following geometric parameters: a diameter of ∼320–460 nm (Figure d,h,l,p) and a depth (H2) of ∼5.6–5.8 μm (Figure c,g,k,o). SEM images demonstrating the technological process of creating silicon tubes based on pillars of various heights: (a)–(d) 5.5 μm; (e)–(h) 4.0 μm; (i)–(l) 2.5 μm; (m)–(p) 1.6 μm. In the micrographs: H1 is the pillar height, H2 is the cavity height. Figure shows the total reflectance spectra in the 200–1600 nm wavelength range for the fabricated tubes and the initial silicon wafer (n-type, (100) orientation, resistivity of 0.004 Ω·cm, single-side polished) used in the etching processes. The reflectance spectrum of the planar silicon sample serves as a reference for comparison and shows good agreement with literature data. , Specifically, in the near UV region (λ = 250–400 nm), two distinct peaks are observed: the E1 peak (∼365 nm or ∼3.4 eV) and the more intense E2 peak (∼272 nm or ∼4.55 eV). The emergence of these maxima is associated with direct interband transitions at critical points in the band structure of crystalline silicon. The E1 and E2 peaks are also present in the reflectance spectra of the fabricated silicon tubes. However, their intensity is reduced compared to the planar sample, which is explained by the influence of the textured surface, leading to enhanced light scattering and local modifications of electronic states. Total reflectance spectra of silicon tube arrays fabricated from pillars of different heights. As can be seen from Figure , the reflectance of the structured samples is significantly lower than that of planar silicon across the entire investigated wavelength range. The spectrally averaged reflectance for the tubes does not exceed ∼18%, whereas for flat silicon, this value is ∼44%. All spectra exhibit a tendency to decrease the reflectance with increasing wavelength (from UV to IR region), while a clear dependence of reflectance on the structure height (H1) is observed. For example, in the 300–1100 nm spectral range, which is key for photovoltaics, the average reflectance is ∼3.8%, ∼7.0%, ∼10.8%, ∼17.2% for structure heights of ∼5.5 μm, ∼4.0 μm, ∼2.5 μm, ∼1.6 μm, respectively. This inverse correlation has been noted in previous studies ,− and can be explained by enhanced light trapping in higher structures. Multiple internal reflections within the spaces between them increase the effective optical path length, leading to greater absorption and reduction in the measured reflectance. In the near IR spectral region, the reflectance of the highest tubes (∼4.0 μm and ∼5.5 μm) remains almost unchanged. For the structures of smaller height (∼1.6 μm and ∼2.5 μm), a slight increase in reflectance is observed at ∼1000 nm, which corresponds to an energy close to the bandgap of silicon (∼1.1 eV). As is known, at wavelengths longer than ∼1100 nm, the photon energy becomes insufficient to excite direct interband transitions in silicon. It results in a sharp decrease in intrinsic absorption and a significant increase in the radiation penetration depth into the material. Consequently, this leads to an increase in the total reflectance due to an additional component arising from reflection from the back surface of the wafer. However, heavily doped silicon does not become transparent even in this spectral region, as the dominant absorption mechanism is replaced by free-carrier absorption. This fact explains the absence of a further increase in reflectance in the IR region in the spectra shown in Figure . Thus, the bands of enhanced reflectance in the 1000–1300 nm range can be attributed to changes in the optical dispersion parameters of the heavily doped silicon near the intrinsic absorption edge. As can be seen, this effect is almost negligible in the structures with greater height, which is most likely related to an increased fraction of absorbed light due to the dominant mechanism of multiple reflections. In other words, such structures act as effective light traps, suppressing its escape back out. To compare the reflectivity of the tubes based on other geometric parameters, namely the height of the inner cavity (H2) and the pillar diameter (Wpillar), additional structures with the following characteristics were fabricated: 1) H1 = ∼2.5 μm, H2 = ∼5.5 μm, Wpillar = ∼0.95 μm (for comparison by pillar diameter, Figure f); 2) H1 = ∼2.7 μm, H2 = ∼4.2 μm, Wpillar = ∼1.35 μm (for comparison by cavity height, Figure d). These structures were individually compared with the previously discussed structures (H1 = ∼2.5 μm, H2 = ∼5.5 μm, Wpillar = ∼1.35 μm) shown in Figure e. (a) Total reflectance spectra of silicon nanotube arrays with different cavity height and different pillar diameters; (b) diffuse reflectance spectrum of a silicon nanotube array with different pillar diameters; (c) specular reflectance spectrum of a silicon nanotube array with different pillar diameters; (d)–(f) SEM images of the corresponding silicon nanotube arrays. As shown in Figure a, a reduction in the cavity height increases the average total reflectance from ∼10.8% to ∼15% across the wavelength range under consideration. However, a smaller pillar diameter enhances the total reflectance only in the UV and visible ranges, while in the IR region, the reflectance for structures with different diameters remains at a similar average level (∼9.3%). This observation can be explained by competing processes that dominate in different spectral ranges, as revealed by analyzing the diffuse and specular reflection components separately. According to the spectra in Figure c, the primary suppression of specular reflection occurs within a range whose upper boundary corresponds to the distance between nearest structures in the array: ∼450–500 nm for structures with a diameter of ∼1.35 μm and ∼850–900 nm for structures with a diameter of ∼0.95 μm. The probable explanation is that at wavelengths exceeding this characteristic distance, the array behaves as a homogeneous medium with a specific effective refractive index gradient, rather than as a collection of individual scattering elements. In this case, reflection is described by Fresnel formulas and determined by the contrast between the effective refractive index of the structure array and that of air. Reducing the structure diameter while maintaining a constant period makes their values closer to each other. As a result, in the IR spectral region, the specular reflectance for structures with a diameter of ∼0.95 μm is significantly lower than for those with a diameter of ∼1.35 μm. When the wavelength becomes comparable to the distance between structures, the key factor becomes the difference in the areas of the flat silicon surface in arrays of different densities: a higher specular reflectance is observed for arrays with a greater distance between tubes. Conversely, the diffuse reflectance is higher for structures with a smaller diameter across the entire studied spectral range, indicating their more efficient light-scattering capability (Figure b). Thus, for structures with smaller pillar diameter, the reduction in the specular component within the IR range is compensated by the decrease in the diffuse component. This ultimately results in similar values for the total reflectance for structures of different diameters. To understand the reflection characteristics in the context of photovoltaic properties, the solar-weighted reflection (R sw ) was calculated using the values of the measured reflectance spectrum and the solar irradiance spectrum (direct + circumsolar) with equation: 9 Rsw=∫λ1λ2R(λ)IAM1.5G(λ)dλ∫λ1λ2IAM1.5G(λ)dλ where R(λ) is the measured total reflection at wavelength λ; I AM1.5G (λ) is the photon flux density of the AM1.5G spectrum at wavelength λ [W·m–2·nm–1]; λ 1 and λ 2 are the boundaries of the integration spectral range. This formula can be expressed in a simplified form since the measurements were performed at discrete wavelengths with a uniform wavelength step: 10 Rsw=∑λ1λ2R(λ)IAM1.5G(λ)∑λ1λ2IAM1.5G(λ)forλ={λmin,λmin+5,...,λmax} The reference values of solar spectral irradiance I AM1.5G (λ) were taken from the ASTM G173-2003 standard table. The calculations resulted in a weighted average reflectance of 3.7%, 6.8%, 9.8%, 16.5% across the 300–1100 nm range for structures with pillar heights of 5.5 μm, 4.0 μm, 2.5 μm, 1.6 μm, respectively. For structures differing in internal cavity height (Figure d), R sw was 13.8%, and for structures differing in pillar width (Figure f), R sw was 11.1%. Measurement of the reflectance of the fabricated structures revealed that the structures with the largest pillar height (∼5.5 μm), the largest cavity height (∼5.8 μm), and the largest pillar diameter (∼1.4 μm) exhibit the lowest reflectivity in the 200–1600 nm wavelength range. The obtained value of ∼3.7% indicates a sufficiently low reflection level compared to analogues described in the literature (Table ). However, a correct comparative analysis must account for the structure geometry, their dimensions, and the parameters of silicon substrate used for their fabrication. 4Comparative Characteristics of Silicon Structures of Various Geometries According to Literature DataRefStructureTotal RWavelength rangePeriod/DensityDiameter/WidthHeightSilicon properties Black silicon0.2%250–400 nm∼6 × 109 cm–2 Dtip = ∼3–5 nm, Dbase = ∼200 nm>5 μmn-type, (100), 0.001–0.005 Ω·cm<1%500–2500 nm Black silicon1.1%300–2500 nm1.7 ± 0.3 pyramids/μm2 -1–2 μmp-type, (100), 0.01–0.02 Ω·cm Black silicon1 ± 0.5%450–1000 nm-150–600 nm3–6 μmn-type, (100), 5–10 Ω·cm Pillars∼3.5%400–1050 nm900 nm400 nm8 μmn-type, (100), 0.06–5 Ω·cm Pillars Rsw∼8.9%300–1100 nm310 nmDbase = ∼270–280 nm, Dtip = ∼20–260 nm∼1.4–1.6 μm- Pillars<10%600–1300 nm460 nm350 nm260 nmn-type, (100) Pillars∼12%450–800 nm400 nm100 nm350 nmn-type Outer Inner Tubes0.9%350–900 nm560 nm380 nm80 nm2.2 μm-Our workTubes3.7%200–1600 nm1.8 μm1.4 μm360 nm5.5 μmn-type, (100), 0.004 Ω·cm Tubes∼4%400–1000 nm1 μm800 nm400 nm3 μmp-type, (001), 1–10 Ω·cm∼26%1000–1600 nm Tubes<8%450–800 nm400 nm200 nm150 nm350 nmn-type Tubes Rsw<10%300–1000 nm500 nm300 nm240 nm1 μm(100) Comparative Characteristics of Silicon Structures of Various Geometries According to Literature Data Conclusion High-aspect-ratio silicon cavities with submicron dimensions and an aspect ratio exceeding 15 were fabricated using Langmuir–Blodgett colloidal lithography, vacuum magnetron sputtering, and continuous plasma chemical etching at “room” temperature of the substrate holder. The etching technology was developed using Bayesian optimization. An initial probabilistic model of the objective function, defines as a superposition of four output characteristics (inverse sidewall angle, undercut, etching isotropy, and aspect ratio), was constructed from two starting points. The algorithm achieved the target values within 12 iterations. Multiple regression analysis, supplemented by optical emission spectroscopy and actinometry, was used to interpret the correlations of silicon plasma etching identified using the Bayesian optimization. The regression model predicted that increasing the C4F8 gas flow rate from 20 to 50 sccm (at a fixed total flow rate) leads to a lower aspect ratio. This is associated with the predominance of the surface passivation mechanism by fluorocarbon compounds over the silicon etching mechanism. This conclusion was confirmed by the results of optical emission spectroscopy and actinometry for C4F8 contents of 60–80% in the total C4F8/SF6 gas mixture. In turn, increasing the pressure from 1.9 to 3.35 Pa results in a larger mask undercut and a greater deviation of the sidewall angle from the vertical. This is attributed to a broadening of the ion angular distribution function and an increase in fluorine radical concentration. A higher ICP power (750 to 1250 W) leads to a greater aspect ratio and mask undercut. This is explained by an increased generation rate of chemically active particles in plasma and a more pronounced enhancement of the vertical etching rate compared to the lateral one. For the same reason (a higher fluorine radical concentration), raising the SF6 flow rate from 7.5 to 22.5 sccm boosts the aspect ratio, while concurrently reducing the etching isotropy and diminishing the inverse sidewall angle. It is important to note that these predictive trends are based on etching in C4F8/SF6 gas mixture, not in a pure SF6 plasma. This means that the silicon surface passivation mechanism is additionally accounted for within the selected ranges of technological parameters. Increasing the bias voltage magnitude from 40 to 150 V leads to a monotonic decrease in the isotropy, mask undercut, and inverse sidewall angle, which is associated with enhanced ion energy and narrowed ion angular distribution function. However, the dependence of the aspect ratio on the bias voltage is nonlinear. This can be explained by the fact that an initial increase in bias voltage magnitude up to 100 V suppresses lateral undercut and amplifies the vertical silicon etching rate. A further increase to 150 V leads to significant erosion of the chromium mask. Since a mask region with sloped walls and reduced thickness is formed at the base if each sphere (near the future etching window edge), intense ion bombardment contributes to faster damage to these areas. This results in an enlarged etching window and a decreased structure aspect ratio. Based on the developed fabrication technology for high-aspect-ratio silicon cavities and previous results on vertically oriented silicon pillars, arrays of monolithic silicon tubes of various heights were fabricated without using standard lithography equipment with an alignment function. The maximum pillar height was 5.5 μm with a cavity depth of 5.8 μm and a diameter of 320 nm (aspect ratio >18). These structures exhibited an average reflection of ∼3.7% across the 200–1600 nm range. It was also found that increasing the pillar height from 1.6 μm to 5.5 μm reduces the average reflection from ∼17.6% to ∼3.7%. Furthermore, decreasing the cavity depth from 5.5 μm to 4.2 μm (for structures with a pillar height of ∼2.5 μm) increases the total reflection across the entire wavelength range. In turn, reducing the pillar diameter from 1.35 μm to 0.95 μm, for structures with a similar cavity geometry and height of ∼5.5 μm, leads to a rise in the total reflection only in the UV and visible ranges. However, in the IR region, the reflectance for structures of different diameters remains at a similar average level (∼9.3%). A comparison of the optical properties of the fabricated tubular structures with works by other authors was conducted. It was established that in the 200–1600 nm wavelength range, the structures in our work exhibit the lowest reflection coefficient (∼3.7%) for pillar aspect ratios below 4. Supplementary Material