- 1.0 A New Product Development Paradigm
- 1.1 Computational Engineering and Virtual Prototypes
- 1.2 Computational Science and Digital Surrogates
- 1.3 The Computational Engineering and Science Ecosystem
- 1.4 High-Performance Computers: The Enablers
- 1.5 Full-Featured Virtual Prototypes
- 1.6 The Advantages of Virtual Prototyping for Systems of Systems
- 1.7 Virtual Prototyping: A Successful Product Development and Scientific Research Paradigm
- 1.8 Historical Perspective
This chapter is from the book
This chapter is from the book
This chapter is from the book
1.2 Computational Science and Digital Surrogates
Science can be defined as follows:
The systematic study of the natural world to identify general laws and principles.
Computational science is defined as follows:
The use of computers, models, and simulations to develop a quantitative understanding of natural phenomena.
Scientific research generally involves four major steps (see Table 1.1):
Table 1.1 Four Major Elements of Scientific Research
1. Observations of natural systems 2. Controlled experiments with the real system or models of the system 3. Development of mathematical theories of the behavior of natural systems 4. Predictions of system behaviors, based on the mathematical theory and machine learning algorithms trained on observations of natural systems |
The first three elements are iterated and can occur in any order. However, predictions require a model and data of sufficient accuracy and breadth. Figure 1.4 illustrates the scientific research process based on these four steps.
)
Figure 1.4 Schematic illustration of the computational science research process (Courtesy Software Engineering Institute)
Computational science has joined experiments and theory as a third leg of the stool that supports scientific advances.
Today observations of natural phenomena such as terrestrial and space weather, planetary systems, chemical reactions, genetic evolution, geologic phenomena, fires, and hundreds of others increasingly rely on the collection and analysis of extremely large data sets. Generally, understanding and predicting the consequences of these natural events require detailed solutions of complex, nonlinear mathematical models that can be obtained only with computers. Additionally, recent advances in general-purpose graphical processing units (GPGPUs) and other accelerator chips optimized for the rapid processing of streaming data flows, together with the development of computational software for multilayer neural networks, have enabled the analysis of very large data sets (many millions of elements) (Krizhevsky 2012 and Rumelhart 1986). Together with the explosion of data from hundreds of different kinds of sensors (cameras, microphones, pressure transducers, magnetism sensors, and so on), computers can be trained using various machine learning algorithms on large sets of data from observations of the natural system and experiments for both real and model systems. This is opening up a revolution in artificial intelligence and can supplement physics-based algorithms.
Investigating the behavior of natural systems focuses on analyzing and developing digital models of natural systems of interest. The models serve as digital surrogates for the actual natural systems. This approach is especially useful for natural systems upon which scientists have little or no influence, such as the weather, climate, and stellar formation and evolution. Scientists can conduct controlled virtual experiments by inserting different physics models into the calculation to see which physics models and sets of initial conditions provide the best match to the observed data. The best physics models can then be used to predict the future behavior of the system. A key advantage of using models for the study of natural systems is that it is often possible to study the details of the behavior of a virtual system’s internal components. This is often very difficult or impossible for many highly interesting real systems (such as supernovae, the interior of the sun, or even large weather systems).
A key difference between computational engineering and computational science is its goals and approaches. One goal of computational engineering is the design of specific, complex products. The software application is a tool for developing the design. All the physics and solution algorithms in the software application are known, verified, validated, and accredited. The outcome of the calculations is the product design. In contrast, the primary goal of computational science is the discovery of knowledge. The purpose of the simulation is to determine which, if any, of a set of candidate physics or other scientific models best fit the observational or experimental data. If none of the models fits the data, new models are needed.
This is the current situation in astrophysics. The universe is now thought to consist of 5% ordinary matter and energy, 27% an unknown type of matter called dark matter (because we can’t see it) (Trimble 1987), and 68% an unknown form of energy known as dark energy, which we also can’t see (Frieman 2008). The evidence for dark matter is this: Dark matter is needed to explain why spiral galaxies don’t fly apart instead of just rotating. There just isn’t enough observed matter to provide the gravitational attraction needed to keep them from flying apart. The evidence for dark energy is that it is needed to explain why the universe is expanding at a rate that continues to increase instead of decreasing as would be expected from gravitational attraction of the observed matter (and dark matter) in the universe. The evidence for the existence of dark matter and dark energy is indirect. It is partly based on the failure of the known physics models to explain the observations described previously. Our lack of any fundamental understanding of 95% of the matter and energy in the universe makes dark matter and dark energy major focal points for current scientific investigation and study.
Weather prediction (meteorology) is a good example of a successful applied computational science research program based on digital surrogates of a complex natural system. The science of weather forecasting began in roughly 1860 with the establishment of the United Kingdom (U.K.) Meteorological Office (MET) by Admiral Robert FitzRoy (Blum 2019). Earlier in his career, FitzRoy was the captain of the HMS Beagle when it carried out a survey of the southern part of South America, with Charles Darwin as the naturalist (1831 to 1836). FitzRoy based his forecasts on weather data sent to him by telegraph from various upwind reporting stations around the U.K., especially those on the western and northern shores. Knowledge of basic principles of meteorology steadily improved during the latter half of the 19th century and the first half of the 20th century (Fleming 2016 and Sawyer 1962), but it still roughly followed FitzRoy’s method of collecting and analyzing upwind weather data (air temperature, air pressure, humidity, wind speed and direction, and precipitation, for example). Weather prediction began to improve more rapidly in the latter half of the 20th century due to the exponential growth of computing following the end of World War II, together with more extensive and more complete data collection and analysis. “Modern 72-hour predictions of hurricane tracks are more accurate than 24-hour forecasts were 40 years ago” (Alley 2019). Alley and his collaborators credit today’s accuracy to improvements in computing power, better understanding of the major physics effects, better data collection, and better computing techniques.
Computational techniques are also proving useful for soft sciences, which involve complex interactions among living systems (animals, people, microbes, and so on) with the natural world. These include biology (medicine, epidemiology, genetics, and others), military strategy and war games, social and political behavior, ecology, economics, finance, business planning, and many other complex systems. Sports teams are even constructing digital twins for their players, to track their physical condition and predict their performance (Siegele 2020).