440
The key advantage of such models is their mechanistic basis, which enhances
their utility to explain system responses to prevailing and counterfactual state
conditions without the need of collecting a complete set of actual data (which
sometimes is impossible to collect). Since these models are based on known
scientific laws, they are typically valid over a wide range of input states and
action choices. However, they can be computationally expensive (e.g., requiring
the solution of partial differential equation systems in high dimensions) and
may require approximations (e.g., linearization) to achieve computational
tractability. These approximations may reduce their accuracy.
To handle accuracy problems and reduce the number of approximations,
supercomputers which have high computing power are widely used for such
tasks. Finally, it should be noted that these models are mostly coded in low- level
languages (e.g., FORTRAN) due to their use of past several decades and the
overhead of higher-level languages (e.g., object-oriented languages) makes them
unattractive. Also, often, these models do not implement multiple optimization
strategies since the models are changing so frequently that hand optimization of
any particular model version is not useful.
Data-driven models are increasingly attractive to the scientists due to the
cheap and abundant data and the availability of technologies that can store,
manage and process such large datasets. These models are capable of capturing
“training examples” of ground truth that can be used to develop empirical relations between the state variables of the system without explicit descriptions of
the underlying physical processes. One example is the identification of farm
fields with known crops and recording of the corresponding satellite observations under a variety of observing conditions and then learning a model that
predicts the crops on the ground from the CCD signals. Similarly, a data-driven
model of hydrology would directly map inputs (precipitation, solar radiation,
and wind) to outputs (evapotranspiration, soil moisture, stream flows).
The key advantage of data-driven models is that in a data-rich world, they
can provide more accurate predictions. Non-parametric, nonlinear models can
adapt to the complexity of the process and yet permit very fast computation of
predictions. Because data-driven models are based on statistical methods, their
uncertainty can be easily quantified. The primary drawback of data-driven
models is that their range of validity is restricted to inputs that are like those
observed during model fitting. Without a mechanistic foundation, the models
have difficulty extrapolating beyond the state conditions that are described by
the “training” data.
FEW systems are rarely stationary because of such factors as climate change,
disturbances, and complex couplings with other systems. This threatens to take
data-driven models outside their “validity horizon” (the region where their predictions are valid). Such data-driven approaches require a large number of datasets, and High-Performance Computing (HPC) infrastructures to avail the
processing. Computational tools that use distributed data structures and analysis (e.g., big data tools such as Hadoop and Spark) are better suited to improve
the I/O (input/output) cost of spatial queries (e.g., retrieving a set of farm
E. Eftelioglu and S. Shekhar
The key advantage of such models is their mechanistic basis, which enhances
their utility to explain system responses to prevailing and counterfactual state
conditions without the need of collecting a complete set of actual data (which
sometimes is impossible to collect). Since these models are based on known
scientific laws, they are typically valid over a wide range of input states and
action choices. However, they can be computationally expensive (e.g., requiring
the solution of partial differential equation systems in high dimensions) and
may require approximations (e.g., linearization) to achieve computational
tractability. These approximations may reduce their accuracy.
To handle accuracy problems and reduce the number of approximations,
supercomputers which have high computing power are widely used for such
tasks. Finally, it should be noted that these models are mostly coded in low- level
languages (e.g., FORTRAN) due to their use of past several decades and the
overhead of higher-level languages (e.g., object-oriented languages) makes them
unattractive. Also, often, these models do not implement multiple optimization
strategies since the models are changing so frequently that hand optimization of
any particular model version is not useful.
Data-driven models are increasingly attractive to the scientists due to the
cheap and abundant data and the availability of technologies that can store,
manage and process such large datasets. These models are capable of capturing
“training examples” of ground truth that can be used to develop empirical relations between the state variables of the system without explicit descriptions of
the underlying physical processes. One example is the identification of farm
fields with known crops and recording of the corresponding satellite observations under a variety of observing conditions and then learning a model that
predicts the crops on the ground from the CCD signals. Similarly, a data-driven
model of hydrology would directly map inputs (precipitation, solar radiation,
and wind) to outputs (evapotranspiration, soil moisture, stream flows).
The key advantage of data-driven models is that in a data-rich world, they
can provide more accurate predictions. Non-parametric, nonlinear models can
adapt to the complexity of the process and yet permit very fast computation of
predictions. Because data-driven models are based on statistical methods, their
uncertainty can be easily quantified. The primary drawback of data-driven
models is that their range of validity is restricted to inputs that are like those
observed during model fitting. Without a mechanistic foundation, the models
have difficulty extrapolating beyond the state conditions that are described by
the “training” data.
FEW systems are rarely stationary because of such factors as climate change,
disturbances, and complex couplings with other systems. This threatens to take
data-driven models outside their “validity horizon” (the region where their predictions are valid). Such data-driven approaches require a large number of datasets, and High-Performance Computing (HPC) infrastructures to avail the
processing. Computational tools that use distributed data structures and analysis (e.g., big data tools such as Hadoop and Spark) are better suited to improve
the I/O (input/output) cost of spatial queries (e.g., retrieving a set of farm
E. Eftelioglu and S. Shekhar