You can use a neural network for this (sort of), and you've got the first step right - train the network to predict $y$ given $x$ (i.e. train it to approximate the function $f$ such that $y = f(x)$). Because the entire neural network is differentiable (presumably), you can take gradients of the input (some $x$) with respect to the predicted output. Then you can use this to update $x$ the same way that you update the weights during training of the network. This let you find a local optimum from $x$. I don't know of a way to try to find a global optimum with a neural network like this except to find the local optimum near many $x$'s and then take the best of those.
If you want a concrete example, take a look at the tutorial on Neural Style Transfer with PyTorch: here you have a noise image as the input and you optimize it to minimize the "style distance" to a reference style image and the "content distance" to a reference content image (i.e. starting with noise, make it look like the content of one image but in the style of another). There's full code there, but here's a short PyTorch snippet that shows the main idea:
# I'm just optimizing one input; clone so that you don't modify the
# original tensor; let it know that we want gradients computed
inputs = train_inputs[0:1].clone().requires_grad_()
input_optimizer = torch.optim.Adam([inputs])
# where `model` is your trained neural network
output = model(inp)
output = 
for _ in range(10000):
output is a list at the end of what the predicted output was after each step of optimizing the input. You can use this to see if you've converged to a local optimum yet (and take more steps/fewer next time as necessary).
inputs will be the optimized input at the end. Note that if there are constraints that your input should satisfy, you'll need to enforce those yourself (e.g. in neural style transfer, they have to enforce that the values are valid for an RGB image).
Note also that how well this works really depends on how well your network approximates $f$; you may well get some unreasonable $x$ for which the network predicts a very large $y$ because there was no training data similar to that $x$ so the network isn't constrained in that region. In general, you should probably be cautious of an $x$ you generate this way that doesn't seem similar to your training data (e.g. has larger/smaller value than any $x$ the network has seen before; interpolation is much easier than extrapolation).
I'm assuming that you have a single data set; if you're able to query for the $y$ values of given $x$'s, then you might want to take a look at Bayesian optimization - essentially the field of trying to find $x$ that maximizes $y = f(x)$ when $f$ is expensive to evaluate and you don't have gradients of it. Bayesian optimization seeks a global optimum.