Appearance
Cutting forces in deep hole drilling cannot be measured directly at the cutting zone. The tool rotates at the bottom of a hole 50–100× its diameter, submerged in high-pressure coolant. Force prediction therefore relies on models — analytical, numerical (FEM), or data-driven (ANN). Choosing the right approach determines whether you can predict tool wear, optimize parameters, or implement real-time monitoring.
Force Components in Deep Hole Drilling
Deep hole drilling forces consist of three distinct components:
| Force Component | Direction | Primary Source | Typical Magnitude (BTA, 40 mm Ø in steel) |
|---|---|---|---|
| Thrust force (Fx) | Axial, along drill axis | Cutting edge engagement | 5–20 kN |
| Torque (Mz) | Rotational | Cutting resistance + friction | 200–800 N·m |
| Radial force | Perpendicular to drill axis | Guide pad contact, imbalance | 1–5 kN |
In BTA drilling, the total torque comprises:
- Cutting torque: Power function of feed rate and diameter
- Friction/chip evacuation torque: Exponential function of hole depth
- Guide pad burnishing torque: Contact friction with the bore wall
Analytical Modeling
Analytical models express forces as functions of cutting parameters using mechanistic equations.
Classical Torque Model
For gun drilling and BTA, torque (M) is typically expressed as:
M = M_c + M_f
Where M_c = C_m · f^a · D^b (cutting torque) And M_f = k · (e^(c·L) − 1) (friction/chip evacuation torque)
| Variable | Meaning |
|---|---|
| f | Feed rate (mm/rev) |
| D | Drill diameter (mm) |
| L | Hole depth (mm) |
| C_m, a, b, k, c | Empirical constants |
Limitations of Analytical Models
| Limitation | Impact |
|---|---|
| Constants are material-specific | Requires calibration for each material |
| Does not capture tool wear effects | Predictions diverge as tool wears |
| Poor generalization to new conditions | Fails when parameters change significantly |
| No depth-dependent dynamics | Average force only |
Artificial Neural Networks (ANN)
ANN models learn the relationship between cutting parameters and forces directly from experimental data, without requiring explicit physical equations.
ANN Architecture for Force Prediction
Chu et al. (2025, Applied Artificial Intelligence) developed a two-layer ANN for torque prediction in ultrasonic-assisted deep hole drilling of AISI-304 stainless steel:
| Network Parameter | Value |
|---|---|
| Architecture | 2 hidden layers |
| Training algorithm | Levenberg-Marquardt |
| Activation function | Hyperbolic tangent (tanh) |
| Inputs | Feed rate, spindle speed, hole depth, vibration amplitude |
| Output | Torque at any depth |
| L/D ratio | 8:1 |
Results vs. nonlinear regression:
| Metric | ANN | Exponential Regression |
|---|---|---|
| Mean absolute percentage error | ~4× smaller | Baseline |
| Standard deviation | ~3.5× smaller | Baseline |
The key innovation was using hole depth as an input vector element rather than treating each depth increment separately, allowing the network to learn the depth-dependent torque characteristic.
ANN for Thrust Force and Torque in Aluminum
A study of AI7075 aluminum drilling compared ANN with response surface methodology (RSM):
| Model | Thrust Force R² | Torque R² |
|---|---|---|
| ANN (two-layer) | 0.9978 | 0.9872 |
| RSM (quadratic) | 0.9314 | 0.9031 |
| FEM simulation | — | Error 6.9% vs. experimental |
Inputs: cutting velocity, feed rate, and tool diameter. ANN significantly outperformed RSM, demonstrating that data-driven methods capture nonlinear interactions that polynomial response surfaces miss.
ANFIS for Tool Condition Monitoring
Sarhan (2025, Arabian Journal for Science and Engineering) developed Adaptive Neuro-Fuzzy Inference System (ANFIS) models as part of an Industry 4.0 tool condition monitoring architecture:
ANFIS model inputs:
- Flank wear (VB)
- Spindle speed
- Feed rate
- Drill diameter
ANFIS model outputs:
- Thrust force
- Torque
ANFIS combines the interpretability of fuzzy logic rules with the learning capability of neural networks. The model enables:
- Real-time force estimation from measured tool wear
- Adaptive feed rate adjustment to maintain constant cutting forces
- Tool change scheduling when predicted forces exceed thresholds
Tip: ANFIS is particularly suitable for production environments where process understanding matters as much as prediction accuracy. Unlike black-box ANN models, ANFIS generates IF-THEN rules that operators can interpret — for example, "IF flank wear is high AND feed rate is high, THEN thrust force is very high."
Hybrid Mechanism-Data-Driven Models
A 2024 study on CFRP drilling combined mechanism-based modeling with neural network compensation:
| Model Type | Mean Relative Error |
|---|---|
| Pure mechanism model | Baseline |
| Hybrid (mechanism + NN compensation) | 1/8 of mechanism (sampled data) |
| Hybrid (mechanism + NN compensation) | 1/3 of mechanism (random data) |
The hybrid approach uses a physical model as the backbone and trains a neural network to predict the residual error — combining the generalization of physics-based models with the accuracy of data-driven fitting.
Finite Element Method (FEM) Simulation
FEM simulation models the cutting process by discretizing the tool-workpiece system into elements and solving the governing equations of deformation, heat transfer, and chip formation.
FEM Software Comparison
| Software | Application in Deep Hole Drilling | Strengths | Limitations |
|---|---|---|---|
| Deform 3D | Chip formation, force prediction, vibration drilling | Specialized for metal forming/machining | Limited tool geometry flexibility |
| ANSYS | Static/dynamic FEA, modal analysis, stress distribution | General-purpose, widely used | Does not model chip formation natively |
| Abaqus | Cutting edge simulation, guide pad burnishing | Advanced contact mechanics, explicit dynamics | Higher setup complexity |
| AdvantEdge | Machining-specific FEM | Pre-configured for orthogonal cutting | Less flexible for non-standard processes |
FEM Results: BTA Drilling of Laminated Materials
A combined analytical-numerical model for BTA drilling of Inconel 625 and FeCr alloy laminated materials (2023, Journal of Manufacturing Processes):
| Predicted Output | Average Error vs. Experiment |
|---|---|
| Torque | 8.4% |
| Feed force | 4.7% |
| Hole expansion | 20.5% |
The model integrated:
- FEM simulation of the cutting edge for primary force prediction
- Analytical modeling of guide pad burnishing for friction forces
- Discrete force summation across multiple cutting edges and pads
FEM Dynamic Simulation
Li et al. used Deform 3D to simulate the dynamic chip formation process in BTA deep hole drilling:
- Torque and axial force increase with drill diameter
- Lower feed rates reduce both force components
- Chip morphology predictions matched experimental results closely
- The FEM model captured the transient force peaks during chip breakage
Modal Analysis for Vibration Avoidance
FEM modal analysis (ANSYS) of the BTA drill stem identifies natural frequencies:
- Highest stress and deformation at the cutting edge and tool block surface
- Drill stem vibration modes are primarily twisting and bending
- Natural frequencies identify dangerous rotational speeds to avoid
- Used for establishing stable spindle speed ranges
Comparative Analysis
Accuracy Comparison
| Method | Thrust Force Error | Torque Error | Training Data Required | Computational Cost |
|---|---|---|---|---|
| Analytical regression | 15–40% | 10–30% | 20–50 experiments | None |
| FEM (Deform/ANSYS) | 5–15% | 7–20% | None (material model) | High (hours per simulation) |
| ANN (data-driven) | 2–8% | 3–10% | 200–1,000+ experiments | Low (seconds to train) |
| ANFIS (hybrid) | 3–10% | 4–12% | 100–500 experiments | Low (minutes to train) |
| Hybrid (FEM + ANN) | 3–7% | 4–8% | 50–200 experiments + FEM | Moderate |
Strengths and Weaknesses
| Method | Strength | Weakness |
|---|---|---|
| Analytical | No experimental data needed; physics-based | Poor accuracy; material constants required |
| FEM | Captures physical phenomena; no training data | Computationally expensive; requires expertise |
| ANN | Highest accuracy; learns nonlinear relationships | Requires large datasets; black-box behaviour |
| ANFIS | Interpretable rules; moderate accuracy | Subjective rule definition; limited complexity |
| Hybrid | Best of both worlds; reduced data needs | Complex to implement; two-stage validation |
Choosing the Right Approach
| Situation | Recommended Approach | Why |
|---|---|---|
| No experimental data, feasibility study | Analytical or FEM | Only options when no data exists |
| Predicting forces for new material | FEM + limited experiments for validation | FEM generalizes to new materials |
| Production monitoring, many data available | ANN or ANFIS | Highest accuracy from existing data |
| Need interpretable rules (quality docs) | ANFIS | IF-THEN rules for process documentation |
| Real-time adaptive control | ANN (lightweight) | Fast inference, embedded deployment |
| Research, understanding mechanisms | FEM | Visualization of stress, strain, chip formation |
| Critical application, high consequence | Hybrid (FEM + ANN) | Redundancy from two independent models |
Implementation Steps
ANN Model Development
- Design of experiments — full factorial or Taguchi matrix covering speed, feed, diameter, depth
- Data collection — measure thrust force and torque using dynamometer at 1–10 kHz sampling
- Preprocessing — filter noise (low-pass at 500 Hz), normalize inputs to [0,1] or [-1,1]
- Architecture selection — start with 2 hidden layers, 10–20 neurons each
- Training — Levenberg-Marquardt or Bayesian regularization; 70/15/15 split for train/val/test
- Validation — compare against independent test set; R² > 0.95 target
- Deployment — export trained network for real-time inference
FEM Model Development
- Geometry creation — 3D model of drill and workpiece (simplified cutting zone)
- Material model — Johnson-Cook constitutive model for plastic deformation
- Meshing — refine mesh at cutting edge (element size 0.01–0.1 mm)
- Boundary conditions — fix workpiece, apply rotation and feed to drill
- Solution — explicit dynamics solver, 0.001–0.01 s simulated cutting time
- Validation — compare thrust and torque with experimental measurements
- Iteration — adjust friction coefficient and material parameters
FAQ
What is the most accurate method for cutting force prediction in deep hole drilling?
ANN models achieve the highest accuracy (R² > 0.99 for thrust force, R² > 0.98 for torque) but require large training datasets. Hybrid FEM-ANN approaches offer the best balance when data is limited.
How does ANN compare to regression for torque prediction?
ANN achieves approximately 4× lower mean absolute percentage error and 3.5× lower standard deviation compared to nonlinear regression (Chu et al., 2025).
What is ANFIS and why is it useful for tool condition monitoring?
ANFIS (Adaptive Neuro-Fuzzy Inference System) combines neural network learning with fuzzy logic interpretability. It produces IF-THEN rules that operators can understand while maintaining good prediction accuracy for thrust force and torque under varying tool wear.
Which FEM software is best for deep hole drilling simulation?
Deform 3D is most commonly used for chip formation and force prediction in drilling. ANSYS is preferred for modal analysis and vibration studies. The choice depends on whether the focus is chip mechanics (Deform) or structural dynamics (ANSYS).
Can cutting forces be predicted for new materials without experimental data?
FEM simulation can predict forces for new materials if the Johnson-Cook material model parameters are available. Accuracy is typically 5–15% error for thrust force and 7–20% for torque.
What is a hybrid mechanism-data-driven model?
A hybrid model uses a physics-based (analytical or FEM) model as a backbone and trains a neural network to predict the residual error. This reduces data requirements while maintaining accuracy — error reduction to 1/8 of pure mechanism models has been demonstrated.
What are the main force components in BTA drilling?
Thrust force (axial), torque (rotational), and radial force. Torque has two sub-components: cutting torque (power function of feed and diameter) and friction/chip evacuation torque (exponential function of hole depth).
How much experimental data is needed for ANN force prediction?
200–1,000 drilling experiments depending on the number of input parameters and desired accuracy. Fewer experiments are needed if a design of experiments (DOE) approach is used.
Can ANFIS be deployed for real-time monitoring?
Yes. ANFIS inference is computationally lightweight and can run on a PLC or edge computing device. Sarhan (2025) proposed an Industry 4.0 architecture with ANFIS for real-time tool condition monitoring.
What is the role of FEM in deep hole drilling research?
FEM enables visualization of stress distribution, chip formation, temperature fields, and tool deformation that cannot be measured experimentally. It is primarily a research and development tool rather than a production monitoring tool.
Conclusion
Cutting force prediction in deep hole drilling is addressed through three complementary approaches. Analytical models provide physics-based estimates without experimental data but with limited accuracy (15–40% error). FEM simulation using Deform 3D or ANSYS enables detailed mechanical analysis of the cutting process with 5–15% error but requires significant computational resources and expertise. ANN models achieve the highest accuracy (R² > 0.99 for thrust force) but require large training datasets and provide limited physical insight. ANFIS offers a practical middle ground with interpretable fuzzy rules and moderate data requirements. Hybrid models combining mechanism-based and data-driven approaches reduce data needs while maintaining accuracy. For production applications, ANFIS or ANN models trained on historical data offer the best combination of accuracy and deployability. For research and development, FEM simulation provides unmeasurable insights into the cutting process mechanics.