
AI in Device Modeling: Redefining Model Accuracy
The semiconductor industry rapidly evolves to more complex design requirements using new materials at novel technology nodes. Artificial neural network (ANN) has surfaced as a game-changing force within this dynamic environment, reshaping how we measure and create device models. This subset of AI can create models without requiring equation development or user-defined parameter extraction, ushering in a new era of innovation in modeling and characterizing semiconductor devices. With growing demand for faster, more accurate models, how can we maximize the potential of ANN? What are the fittest use cases?
What is ANN?
An Artificial Neural Network (ANN) is a computational model inspired by the structure and functioning of the human brain. ANNs consist of layers of artificial neurons, also known as nodes or units. Neurons within one layer establish connections, characterized by weights and biases, with others in the adjacent layer, forging a dynamic network. ANNs learn by adjusting their weights and biases iteratively to minimize the error between predicted and actual outputs. This process of forward and backward passes, known as feedforward and backpropagation, continues until the network converges to a desired level of accuracy.
Figure 1. The basics of Artificial Neural Networks
Conceptually, it only requires three steps: measure, train, and simulate. No need for complex equations or user-defined parameters. It all boils down to one key factor – good data with the right training methods to fit the simulation to the measured data.
4 key reasons to leverage ANN in device modeling
Artificial Intelligence (AI) presents a wealth of advantages for device modeling teams. Here are four crucial scenarios where we strongly advocate the utilization of ANNs:
Rapid model generation
ANNs shine when there's a need for quick model creation at any stage of the design process. ANNs are capable of swiftly generating highly accurate models that closely align with measured data.
Non-linear model fitting
In cases where a standard model isn't readily available or when dealing with cutting-edge technologies, ANNs excel at fitting non-linear models to measured data, when the compact model is not accurate enough.
Accelerated extraction
ANNs allow for ease of extraction for some large parameter compact models. As exemplified by a Keysight Labs modeling engineer, the transition from a four-month modeling process to a mere two minutes for a GaS (Gallium Selenide) device underscores the efficiency ANNs can bring to extraction tasks.
Intellectual property protection
For those concerned with safeguarding their intellectual property (IP), ANNs provide a valuable solution to generate intricate models efficiently while protecting proprietary designs.
Incorporating ANNs into device modeling processes not only expedites model generation but also empowers teams to handle the complexities in modeling novel devices. But it's important to know when not to use them. For instance, if you have a well-working model like BSIM 4 or need to fine-tune an existing one, ANNs might not be the best choice. Deciding when to use ANNs should involve careful consideration of your modeling needs and the effectiveness of your current models.
What sets Keysight’s ANN modeling solution apart?
Figure 2. IC-CAP ANN Modeling Toolkit
Our journey with Artificial Neural Networks (ANNs) began as an internal endeavor by the Keysight Lab. We recognized the need for cutting-edge tools to support our quest for the fastest and broadest bandwidth spectrum analyzers and oscilloscopes in the market. This internal genesis ensures that our ANN implementation is not just powerful but tailored to meet the unique demands of the semiconductor industry.
Moreover, unlike many popular generic ANNs, our ANN modeling solution is designed with a strong focus on device modeling. This specialized approach enables us to address the intricate nuances and needs, for example, the capability to handle partial derivative data as charge from capacitance.
Thirdly, it outputs a standard Verilog-A model. This output format offers great flexibility, allowing users to seamlessly integrate the models into SPICE simulations, enhancing the overall modeling and simulation efficiency.
Final words
As we look ahead, in the ever-changing realm of semiconductor devices, the rise of AI will continue to push the disruptive shift in device modeling and simulation. Choosing a pioneering ANN modeling solution empowers modeling teams to unlock the immense possibilities this technology offers and navigate the dynamic landscape with greater resilience.